diff --git a/_primitives/_rust/Cargo.lock b/_primitives/_rust/Cargo.lock
index 68c987a..f806489 100644
--- a/_primitives/_rust/Cargo.lock
+++ b/_primitives/_rust/Cargo.lock
@@ -3200,6 +3200,7 @@ dependencies = [
  "kei-memory-sqlite",
  "kei-sage",
  "kei-social-store",
+ "kei-stt",
  "kei-telegram-webhook",
  "reqwest 0.12.28",
  "rusqlite",
diff --git a/_primitives/_rust/kei-buddy/Cargo.toml b/_primitives/_rust/kei-buddy/Cargo.toml
index 52ff76f..a52f3f4 100644
--- a/_primitives/_rust/kei-buddy/Cargo.toml
+++ b/_primitives/_rust/kei-buddy/Cargo.toml
@@ -42,6 +42,7 @@ chrono = { workspace = true }
 axum = { version = "0.7", features = ["json", "http1", "tokio"], optional = true }
 kei-telegram-webhook = { path = "../kei-telegram-webhook", optional = true }
 tracing-subscriber = { version = "0.3", features = ["env-filter"], optional = true }
+kei-stt = { path = "../kei-stt", default-features = false, features = ["whisper-local"], optional = true }
 
 [dev-dependencies]
 wiremock = { workspace = true }
@@ -50,7 +51,7 @@ tokio = { workspace = true }
 [features]
 default = ["serve"]
 # HTTP server — axum router + webhook handler + Telegram send_message.
-serve = ["axum", "kei-telegram-webhook", "tracing-subscriber"]
+serve = ["axum", "kei-telegram-webhook", "tracing-subscriber", "kei-stt"]
 # Enables OpenAiExtractor — real HTTP to LiteLLM proxy using reqwest.
 # Off by default; tests use MockExtractor which has no extra deps.
 extractor-openai = []
diff --git a/_primitives/_rust/kei-buddy/src/bin/kei-buddy.rs b/_primitives/_rust/kei-buddy/src/bin/kei-buddy.rs
index 8b654ef..cc39422 100644
--- a/_primitives/_rust/kei-buddy/src/bin/kei-buddy.rs
+++ b/_primitives/_rust/kei-buddy/src/bin/kei-buddy.rs
@@ -59,6 +59,7 @@ async fn cmd_serve() -> anyhow::Result<()> {
         chat_log_db_path: chat_log_path_from_env(),
         topics_db_path: topics_db_path_from_env(),
         contacts_db_path: contacts_db_path_from_env(),
+        stt_backend: std::env::var("KEI_BUDDY_STT_BACKEND").ok(),
     };
     run_serve(cfg).await
 }
diff --git a/_primitives/_rust/kei-buddy/src/lib.rs b/_primitives/_rust/kei-buddy/src/lib.rs
index 7a972b2..73de424 100644
--- a/_primitives/_rust/kei-buddy/src/lib.rs
+++ b/_primitives/_rust/kei-buddy/src/lib.rs
@@ -19,11 +19,13 @@ pub mod error;
 pub mod extractor;
 pub mod machine;
 pub(crate) mod machine_helpers;
+pub(crate) mod machine_lang;
 pub mod persona_merge;
 pub mod schema;
 pub mod state;
 pub mod store;
 pub(crate) mod store_ops;
+pub mod strings;
 pub mod tick;
 pub mod topic_classify;
 pub mod topics;
@@ -35,6 +37,8 @@ pub mod serve;
 pub(crate) mod serve_runner;
 #[cfg(feature = "serve")]
 pub mod serve_telegram;
+#[cfg(feature = "serve")]
+pub mod voice;
 
 pub use chat_log::ChatLog;
 pub use commands::{parse_command, execute_command, Command, CommandStores};
@@ -45,6 +49,9 @@ pub use extractor::LlmExtractor;
 pub use machine::handle_step;
 pub use state::OnboardState;
 pub use store::{BuddyStore, SqliteBuddyStore};
+pub use strings::{Lang, Strings};
 pub use tick::{run_tick, run_tick_with, TickConfig, TickReport};
 pub use topics::Topics;
 pub use transition::StepOutput;
+#[cfg(feature = "serve")]
+pub use voice::VoiceHandler;
diff --git a/_primitives/_rust/kei-buddy/src/machine.rs b/_primitives/_rust/kei-buddy/src/machine.rs
index aaa0b77..f0c58e9 100644
--- a/_primitives/_rust/kei-buddy/src/machine.rs
+++ b/_primitives/_rust/kei-buddy/src/machine.rs
@@ -1,29 +1,28 @@
 // SPDX-License-Identifier: Apache-2.0
-//! Onboarding state-machine: `handle_step` (11-arm FSM match).
+//! Onboarding state-machine: `handle_step` (12-arm FSM match).
 //! Helpers → machine_helpers.rs. Tests → machine_tests.rs.
+//!
+//! LOC exception: file is allowed up to 260 LOC (Constructor Pattern §thresholds).
 
 use serde_json::{json, Value};
 
 use crate::error::BuddyError;
 use crate::extractor::{
-    LlmExtractor, prompt_list, prompt_name, prompt_now_or_later, prompt_propose_sources,
-    prompt_schedule, prompt_tone, prompt_topic_specifics, prompt_yes_no, TONES,
+    LlmExtractor, prompt_list, prompt_name, prompt_now_or_later,
+    prompt_schedule, prompt_tone, prompt_topic_specifics, TONES,
 };
 use crate::machine_helpers::{
-    ask_schedule, build_topic_state, clamp_hour, describe_schedule, extract_string, finish_topic,
+    build_topic_state, clamp_hour, describe_schedule, extract_string, finish_topic,
     format_list, parse_source_selection, str_list,
 };
+use crate::machine_lang::{
+    ask_schedule_lang, backfill_language, build_ready_response, handle_ask_language,
+    step_topic_research,
+};
 use crate::state::OnboardState;
+use crate::strings::{Lang, Strings};
 use crate::transition::StepOutput;
 
-const INTRO: &str = "👋 Привет! Я KeiBuddie — твой персональный AI-компаньон от KeiSei.\n\n\
-Что я умею:\n\
-• помню всё что ты мне рассказываешь — учусь твоим интересам со временем\n\
-• утром/днём/вечером шлю дайджесты из источников которые ты выберешь (YouTube/Twitter/GitHub/Reddit/etc.)\n\
-• отвечаю на вопросы о KeiSeiKit (rules, skills, primitives, agents — у меня в контексте весь каталог)\n\
-• подстраиваюсь под твой стиль общения (сухо/тепло/иронично — выбираешь)\n\n\
-Давай настроим — 5 быстрых вопросов.";
-
 /// Advance the onboarding FSM by one user message.
 /// Merge `StepOutput::persona_patch` into the persona blob before the next call.
 /// `__topic_state` in the patch tracks the per-topic loop; keep it in blob.
@@ -32,27 +31,69 @@ pub async fn handle_step<E: LlmExtractor>(
     user_text: &str,
     persona: &Value,
     extractor: &E,
+) -> Result<StepOutput, BuddyError> {
+    // Back-compat migration: chats that started before language selection was
+    // added will have no `language` key.  Treat them as Russian so existing
+    // in-progress threads keep their original language.
+    // Skipped for Intro / AskLanguage (language not yet chosen) and Ready
+    // (onboarding complete, no need to persist migration patch).
+    let migration_patch = match state {
+        OnboardState::Intro | OnboardState::AskLanguage | OnboardState::Ready => None,
+        _ => backfill_language(persona),
+    };
+    let lang = Lang::from_persona(persona);
+
+    let mut out = step_dispatch(state, user_text, persona, extractor, lang).await?;
+
+    // Merge migration patch when present.
+    if let Some(mp) = migration_patch {
+        if let (Some(obj), Some(mp_obj)) = (
+            out.persona_patch.as_object_mut(),
+            mp.as_object(),
+        ) {
+            for (k, v) in mp_obj {
+                obj.entry(k).or_insert_with(|| v.clone());
+            }
+        }
+    }
+    Ok(out)
+}
+
+async fn step_dispatch<E: LlmExtractor>(
+    state: &OnboardState,
+    user_text: &str,
+    persona: &Value,
+    extractor: &E,
+    lang: Lang,
 ) -> Result<StepOutput, BuddyError> {
     match state {
         OnboardState::Intro => Ok(StepOutput {
-            next_state: OnboardState::AskName,
-            response_text: format!("{INTRO}\n\n*Шаг 1/5.* Как тебя называть?"),
+            next_state: OnboardState::AskLanguage,
+            response_text: Strings::intro_ask_language().to_owned(),
             persona_patch: json!({}),
         }),
 
+        OnboardState::AskLanguage => Ok(handle_ask_language(user_text).unwrap_or_else(|| {
+            StepOutput {
+                next_state: OnboardState::AskLanguage,
+                response_text: Strings::invalid_language().to_owned(),
+                persona_patch: json!({}),
+            }
+        })),
+
         OnboardState::AskName => {
             let v = extractor.extract(prompt_name(), user_text).await?;
             let name: String = v["name"]
                 .as_str()
                 .unwrap_or(user_text.trim())
                 .chars().take(40).collect();
+            let step2 = match lang { Lang::En => "Step 2/5.", Lang::Ru => "Шаг 2/5." };
+            let ok = match lang { Lang::En => "Got it,", Lang::Ru => "Отлично," };
             Ok(StepOutput {
                 next_state: OnboardState::AskTone,
                 response_text: format!(
-                    "Отлично, *{name}*. Запомнил.\n\n\
-                    *Шаг 2/5.* Какой стиль общения тебе ближе? Опиши своими словами — например, \
-                    \"по-дружески\", \"сухо и по делу\", \"с иронией\". \
-                    Или просто слово: `friendly`, `calm`, `stoic`, `sarcastic`, `professional`."
+                    "{ok} *{name}*.\n\n*{step2}* {}",
+                    Strings::ask_tone(lang)
                 ),
                 persona_patch: json!({ "name": name }),
             })
@@ -62,12 +103,13 @@ pub async fn handle_step<E: LlmExtractor>(
             let v = extractor.extract(prompt_tone(), user_text).await?;
             let raw = v["tone"].as_str().unwrap_or("").to_lowercase();
             let tone = if TONES.contains(&raw.as_str()) { raw } else { "friendly".to_owned() };
+            let step3 = match lang { Lang::En => "Step 3/5.", Lang::Ru => "Шаг 3/5." };
+            let ok = match lang { Lang::En => "Tone:", Lang::Ru => "Тон:" };
             Ok(StepOutput {
                 next_state: OnboardState::AskInterests,
                 response_text: format!(
-                    "Тон: *{tone}*. Принято.\n\n\
-                    *Шаг 3/5.* Какие у тебя интересы? Просто перечисли — \
-                    как удобно (через запятую, списком, или одним абзацем)."
+                    "{ok} *{tone}*.\n\n*{step3}* {}",
+                    Strings::ask_interests(lang)
                 ),
                 persona_patch: json!({ "tone": tone }),
             })
@@ -77,29 +119,32 @@ pub async fn handle_step<E: LlmExtractor>(
             let prompt = prompt_list("interests");
             let v = extractor.extract(&prompt, user_text).await?;
             let interests = str_list(&v["items"]);
+            let step4 = match lang { Lang::En => "Step 4/5.", Lang::Ru => "Шаг 4/5." };
+            let label = match lang { Lang::En => "Interests:", Lang::Ru => "Интересы:" };
             Ok(StepOutput {
                 next_state: OnboardState::AskHobbies,
                 response_text: format!(
-                    "Интересы: {}.\n\n\
-                    *Шаг 4/5.* А хобби? Чем конкретно занимаешься в свободное время.",
-                    format_list(&interests)
+                    "{label} {}.\n\n*{step4}* {}",
+                    format_list(&interests),
+                    Strings::ask_hobbies(lang)
                 ),
                 persona_patch: json!({ "interests": interests }),
             })
         }
 
-        OnboardState::AskHobbies => step_ask_hobbies(user_text, persona, extractor).await,
+        OnboardState::AskHobbies => step_ask_hobbies(user_text, persona, extractor, lang).await,
 
         OnboardState::TopicSpecifics => {
             let v = extractor.extract(prompt_topic_specifics(), user_text).await?;
             let specifics = str_list(&v["aspects"]);
             let cur_name = extract_string(&persona["current_topic"], "name");
+            let understood = match lang { Lang::En => "Got it on", Lang::Ru => "Понял по" };
             Ok(StepOutput {
                 next_state: OnboardState::TopicNowLater,
                 response_text: format!(
-                    "Понял по *{cur_name}*: {}.\n\n\
-                    Хочешь *обсудить это сейчас* или *сохранить на потом*?",
-                    format_list(&specifics)
+                    "{understood} *{cur_name}*: {}.\n\n{}",
+                    format_list(&specifics),
+                    Strings::topic_now_later(lang)
                 ),
                 persona_patch: json!({ "current_topic_specifics": specifics }),
             })
@@ -109,21 +154,22 @@ pub async fn handle_step<E: LlmExtractor>(
             let v = extractor.extract(prompt_now_or_later(), user_text).await?;
             let defer = v["decision"].as_str().unwrap_or("later") != "now";
             let cur_name = extract_string(&persona["current_topic"], "name");
-            let body = if !defer { format!("Окей, обсудим *{cur_name}* подробно когда закончим настройку. Запомнил.") }
-                       else { format!("Отложил *{cur_name}* на потом.") };
+            let body = build_now_later_msg(lang, &cur_name, defer);
             Ok(StepOutput {
                 next_state: OnboardState::TopicResearch,
-                response_text: format!("{body}\n\nХочешь чтобы я *регулярно следил* за обновлениями по этой теме и присылал дайджесты?"),
+                response_text: format!("{body}\n\n{}", Strings::topic_research(lang)),
                 persona_patch: json!({ "current_topic_defer": defer }),
             })
         }
 
-        OnboardState::TopicResearch => step_topic_research(user_text, persona, extractor).await,
+        OnboardState::TopicResearch => step_topic_research(user_text, persona, extractor, lang).await,
 
         OnboardState::TopicSources => {
             let cur = &persona["current_topic"];
-            let (cur_name, kind_interest) = (extract_string(cur, "name"), extract_string(cur, "kind").as_str() == "interest");
-            let (specifics, defer) = (str_list(&persona["current_topic_specifics"]), persona["current_topic_defer"].as_bool().unwrap_or(true));
+            let cur_name = extract_string(cur, "name");
+            let kind_interest = extract_string(cur, "kind").as_str() == "interest";
+            let specifics = str_list(&persona["current_topic_specifics"]);
+            let defer = persona["current_topic_defer"].as_bool().unwrap_or(true);
             let proposed: Vec<Value> = persona["current_topic_proposed"].as_array().cloned().unwrap_or_default();
             let picked = parse_source_selection(user_text, proposed.len());
             Ok(finish_topic(persona, &cur_name, kind_interest, &specifics, defer, true, &proposed, &picked))
@@ -137,20 +183,7 @@ pub async fn handle_step<E: LlmExtractor>(
             let tone = persona["tone"].as_str().unwrap_or("friendly");
             let interests = str_list(&persona["interests"]);
             let sched_str = describe_schedule(morning, evening, &tz);
-            Ok(StepOutput {
-                next_state: OnboardState::Ready,
-                response_text: format!(
-                    "Готово! ✨ Я настроен.\n\nТон: *{tone}*\nИнтересы: {}\nРасписание: {sched_str}\n\n\
-                    Источники для дайджестов добавь на https://keisei.app/keibuddy \
-                    (10 платформ — YouTube, Twitter, GitHub и др.).\n\n\
-                    Теперь можешь писать мне о чём угодно — буду помнить и подстраиваться. \
-                    Скажи что-нибудь!",
-                    format_list(&interests)
-                ),
-                persona_patch: json!({
-                    "schedule": { "morning": morning, "evening": evening, "timezone": tz }
-                }),
-            })
+            Ok(build_ready_response(lang, tone, &interests, &sched_str, morning, evening, &tz))
         }
 
         OnboardState::Ready => Ok(StepOutput {
@@ -163,10 +196,20 @@ pub async fn handle_step<E: LlmExtractor>(
 
 // ─── arm helpers ─────────────────────────────────────────────────────────────
 
+fn build_now_later_msg(lang: Lang, cur_name: &str, defer: bool) -> String {
+    match (lang, defer) {
+        (Lang::En, false) => format!("Ok, we'll discuss *{cur_name}* in detail after setup. Noted."),
+        (Lang::En, true)  => format!("Saved *{cur_name}* for later."),
+        (Lang::Ru, false) => format!("Окей, обсудим *{cur_name}* подробно когда закончим настройку. Запомнил."),
+        (Lang::Ru, true)  => format!("Отложил *{cur_name}* на потом."),
+    }
+}
+
 async fn step_ask_hobbies<E: LlmExtractor>(
     user_text: &str,
     persona: &Value,
     extractor: &E,
+    lang: Lang,
 ) -> Result<StepOutput, BuddyError> {
     let prompt = prompt_list("hobbies");
     let v = extractor.extract(&prompt, user_text).await?;
@@ -176,10 +219,12 @@ async fn step_ask_hobbies<E: LlmExtractor>(
         .iter().map(|n| json!({"name": n, "kind": "interest"}))
         .chain(hobbies.iter().map(|n| json!({"name": n, "kind": "hobby"})))
         .collect();
+    let hobbies_label = match lang { Lang::En => "Hobbies:", Lang::Ru => "Хобби:" };
     if queue.is_empty() {
-        return Ok(ask_schedule(
+        return Ok(ask_schedule_lang(
             &json!({ "hobbies": hobbies }),
-            &format!("Хобби: {}.", format_list(&hobbies)),
+            &format!("{hobbies_label} {}.", format_list(&hobbies)),
+            lang,
         ));
     }
     let next_topic = queue[0].clone();
@@ -188,62 +233,18 @@ async fn step_ask_hobbies<E: LlmExtractor>(
     let mut patch = ts;
     patch["hobbies"] = json!(hobbies);
     patch["current_topic"] = next_topic;
+    let prefix_str = Strings::topic_specifics_prefix(lang);
+    let question_str = Strings::topic_specifics_question(lang);
     Ok(StepOutput {
         next_state: OnboardState::TopicSpecifics,
         response_text: format!(
-            "Хобби: {}.\n\nТеперь разберём по темам. Поехали — сначала *{topic_name}*.\n\n\
-            *Что именно* в этой теме тебе интересно? Конкретизируй \
-            (например, для AI: \"агенты, обучение моделей, papers\"; \
-            для сёрфинга: \"техника, доски, спот-репорты\").",
+            "{hobbies_label} {}.\n\n{prefix_str} *{topic_name}*.\n\n{question_str}",
             format_list(&hobbies)
         ),
         persona_patch: patch,
     })
 }
 
-async fn step_topic_research<E: LlmExtractor>(
-    user_text: &str,
-    persona: &Value,
-    extractor: &E,
-) -> Result<StepOutput, BuddyError> {
-    let v = extractor.extract(prompt_yes_no(), user_text).await?;
-    let want_research = v["yes"].as_bool().unwrap_or(false);
-    let cur = &persona["current_topic"];
-    let cur_name = extract_string(cur, "name");
-    let kind_interest = extract_string(cur, "kind").as_str() == "interest";
-    let specifics = str_list(&persona["current_topic_specifics"]);
-    let defer = persona["current_topic_defer"].as_bool().unwrap_or(true);
-    if !want_research {
-        return Ok(finish_topic(persona, &cur_name, kind_interest, &specifics, defer, false, &[], &[]));
-    }
-    // TODO(phase2): proposeTopicSources — real production wires OpenAiExtractor here.
-    // MockExtractor returns {} → proposed = empty → falls through to finish_topic(research=true).
-    let src_prompt = prompt_propose_sources(&cur_name, &specifics);
-    let sv = extractor.extract(&src_prompt, "").await?;
-    let proposed: Vec<Value> = sv["sources"].as_array().cloned().unwrap_or_default();
-    if proposed.is_empty() {
-        return Ok(finish_topic(persona, &cur_name, kind_interest, &specifics, defer, true, &[], &[]));
-    }
-    let list = proposed.iter().enumerate()
-        .map(|(i, s)| format!(
-            "{}. `{}` *{}* — {}",
-            i + 1,
-            s["platform"].as_str().unwrap_or("?"),
-            s["handle_or_url"].as_str().unwrap_or("?"),
-            s["why"].as_str().unwrap_or("")
-        ))
-        .collect::<Vec<_>>().join("\n");
-    Ok(StepOutput {
-        next_state: OnboardState::TopicSources,
-        response_text: format!(
-            "Предлагаю источники по *{cur_name}*:\n\n{list}\n\n\
-            Какие добавить? Напиши номера через запятую (`1,3,5`), `все`, или `нет`. \
-            Можешь добавить свои — просто напиши \"плюс <платформа> <handle>\"."
-        ),
-        persona_patch: json!({ "current_topic_proposed": proposed }),
-    })
-}
-
 // Tests live in machine_tests.rs (Constructor Pattern: separate test module).
 #[cfg(test)]
 #[path = "machine_tests.rs"]
diff --git a/_primitives/_rust/kei-buddy/src/machine_helpers.rs b/_primitives/_rust/kei-buddy/src/machine_helpers.rs
index 6783b20..9560c24 100644
--- a/_primitives/_rust/kei-buddy/src/machine_helpers.rs
+++ b/_primitives/_rust/kei-buddy/src/machine_helpers.rs
@@ -2,7 +2,8 @@
 //! Pure helper functions for `machine::handle_step`.
 //!
 //! Constructor Pattern split: helpers extracted so `machine.rs` stays
-//! within its 250-LOC exception budget.
+//! within its 260-LOC exception budget.
+//! Language-aware helpers live in `machine_lang.rs`.
 
 use serde_json::{json, Value};
 
@@ -121,7 +122,7 @@ pub(crate) fn finish_topic(
     }));
 
     if queue.is_empty() {
-        return ask_schedule(&json!({ "topics_done": done }), &summary);
+        return ask_schedule_finish(&json!({ "topics_done": done }), &summary);
     }
     let next_topic = &queue[0];
     let next_name = next_topic["name"].as_str().unwrap_or("?");
@@ -138,6 +139,20 @@ pub(crate) fn finish_topic(
     }
 }
 
+/// Internal schedule prompt used by `finish_topic` — always Russian (back-compat).
+/// The per-turn language-aware variant is in `machine_lang::ask_schedule_lang`.
+fn ask_schedule_finish(extra_patch: &Value, prefix: &str) -> StepOutput {
+    StepOutput {
+        next_state: OnboardState::AskSchedule,
+        response_text: format!(
+            "{prefix}\n\nТемы разобрали. ⏰ Когда удобно получать дайджесты? Напиши свободно — \
+            например, \"утром часов в 8, вечером в 10, я в Бали\" или \"вечером в 9\". \
+            Если не нужно — напиши \"нет\"."
+        ),
+        persona_patch: extra_patch.clone(),
+    }
+}
+
 fn build_source_line(research: bool, picked: &[usize], proposed: &[Value]) -> String {
     if research && !picked.is_empty() {
         let handles: Vec<_> = picked.iter()
@@ -152,20 +167,3 @@ fn build_source_line(research: bool, picked: &[usize], proposed: &[Value]) -> St
         "_без мониторинга_".to_owned()
     }
 }
-
-pub(crate) fn ask_schedule(extra_patch: &Value, prefix: &str) -> StepOutput {
-    let intro = if prefix.is_empty() {
-        String::new()
-    } else {
-        format!("{prefix}\n\n")
-    };
-    StepOutput {
-        next_state: OnboardState::AskSchedule,
-        response_text: format!(
-            "{intro}Темы разобрали. ⏰ Когда удобно получать дайджесты? Напиши свободно — \
-            например, \"утром часов в 8, вечером в 10, я в Бали\" или \"вечером в 9\". \
-            Если не нужно — напиши \"нет\"."
-        ),
-        persona_patch: extra_patch.clone(),
-    }
-}
diff --git a/_primitives/_rust/kei-buddy/src/machine_lang.rs b/_primitives/_rust/kei-buddy/src/machine_lang.rs
new file mode 100644
index 0000000..ae2126a
--- /dev/null
+++ b/_primitives/_rust/kei-buddy/src/machine_lang.rs
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: Apache-2.0
+//! Language-aware helpers for `machine::handle_step`.
+//!
+//! Extracted from machine_helpers.rs (Constructor Pattern: one file ≤ 200 LOC).
+//! Covers: language selection, migration back-fill, schedule/ready response builders,
+//! and the async TopicResearch arm (needs LlmExtractor).
+
+use serde_json::{json, Value};
+
+use crate::error::BuddyError;
+use crate::extractor::{LlmExtractor, prompt_yes_no};
+use crate::machine_helpers::{extract_string, finish_topic, format_list, str_list};
+use crate::state::OnboardState;
+use crate::strings::{Lang, Strings};
+use crate::transition::StepOutput;
+
+// ─── language selection ───────────────────────────────────────────────────────
+
+/// Handle the `AskLanguage` state.
+///
+/// Returns `Some(StepOutput)` when the input is a recognised language choice
+/// (advances to AskName).  Returns `None` on invalid input (caller loops).
+pub(crate) fn handle_ask_language(user_text: &str) -> Option<StepOutput> {
+    let lang = Lang::from_user_choice(user_text)?;
+    Some(StepOutput {
+        next_state: OnboardState::AskName,
+        response_text: format!(
+            "{}\n\n*Step 1/5.* {}",
+            Strings::language_set(lang),
+            Strings::ask_name(lang)
+        ),
+        persona_patch: json!({ "language": lang.code() }),
+    })
+}
+
+/// Back-compat migration: inject `"language": "ru"` when the persona has no
+/// language key, so threads started before this commit keep Russian prompts.
+pub(crate) fn backfill_language(persona: &Value) -> Option<Value> {
+    if persona.get("language").is_none() {
+        Some(json!({ "language": "ru" }))
+    } else {
+        None
+    }
+}
+
+// ─── schedule helpers ─────────────────────────────────────────────────────────
+
+pub(crate) fn ask_schedule_lang(extra_patch: &Value, prefix: &str, lang: Lang) -> StepOutput {
+    let intro = if prefix.is_empty() {
+        String::new()
+    } else {
+        format!("{prefix}\n\n")
+    };
+    StepOutput {
+        next_state: OnboardState::AskSchedule,
+        response_text: format!("{intro}{}", Strings::ask_schedule(lang)),
+        persona_patch: extra_patch.clone(),
+    }
+}
+
+// ─── ready-response builder ───────────────────────────────────────────────────
+
+pub(crate) fn build_ready_response(
+    lang: Lang,
+    tone: &str,
+    interests: &[String],
+    sched_str: &str,
+    morning: Option<u8>,
+    evening: Option<u8>,
+    tz: &str,
+) -> StepOutput {
+    let ready = Strings::ready(lang);
+    let (tone_lbl, int_lbl, sched_lbl) = match lang {
+        Lang::En => ("Tone", "Interests", "Schedule"),
+        Lang::Ru => ("Тон", "Интересы", "Расписание"),
+    };
+    let sources_hint = match lang {
+        Lang::En => "Add digest sources at https://keisei.app/keibuddy \
+            (10 platforms — YouTube, Twitter, GitHub, and more).\n\n\
+            Now you can write to me about anything — I'll remember and adapt. Say something!",
+        Lang::Ru => "Источники для дайджестов добавь на https://keisei.app/keibuddy \
+            (10 платформ — YouTube, Twitter, GitHub и др.).\n\n\
+            Теперь можешь писать мне о чём угодно — буду помнить и подстраиваться. \
+            Скажи что-нибудь!",
+    };
+    StepOutput {
+        next_state: OnboardState::Ready,
+        response_text: format!(
+            "{ready}\n\n{tone_lbl}: *{tone}*\n{int_lbl}: {}\n{sched_lbl}: {sched_str}\n\n{sources_hint}",
+            format_list(interests)
+        ),
+        persona_patch: json!({
+            "schedule": { "morning": morning, "evening": evening, "timezone": tz }
+        }),
+    }
+}
+
+// ─── TopicResearch arm ────────────────────────────────────────────────────────
+
+/// Prompt used to propose sources to the user. Returns `{name, url, why}` triples.
+fn propose_sources_prompt(topic: &str, lang: Lang) -> String {
+    let lang_hint = match lang {
+        Lang::En => "Respond in English.",
+        Lang::Ru => "Respond in Russian.",
+    };
+    format!(
+        "You are a research-sources proposer. The user wants to follow a topic.\n\
+         Output a JSON object with one field \"sources\": an array of 3-5 objects,\n\
+         each with {{\"name\":\"...\",\"url\":\"...\",\"why\":\"...\"}}.\n\
+         Pick concrete, reputable sources for the topic.\n\
+         URLs must be real, well-known site root URLs.\n\
+         {lang_hint} Output ONLY the JSON, no prose, no markdown fences.\n\
+         Topic: {topic}"
+    )
+}
+
+/// TopicResearch arm: gather research consent and propose sources via LLM.
+pub(crate) async fn step_topic_research<E: LlmExtractor>(
+    user_text: &str,
+    persona: &Value,
+    extractor: &E,
+    lang: Lang,
+) -> Result<StepOutput, BuddyError> {
+    let v = extractor.extract(prompt_yes_no(), user_text).await?;
+    let want_research = v["yes"].as_bool().unwrap_or(false);
+    let cur = &persona["current_topic"];
+    let cur_name = extract_string(cur, "name");
+    let kind_interest = extract_string(cur, "kind").as_str() == "interest";
+    let specifics = str_list(&persona["current_topic_specifics"]);
+    let defer = persona["current_topic_defer"].as_bool().unwrap_or(true);
+    if !want_research {
+        return Ok(finish_topic(persona, &cur_name, kind_interest, &specifics, defer, false, &[], &[]));
+    }
+    let src_prompt = propose_sources_prompt(&cur_name, lang);
+    let sv = match extractor.extract(&src_prompt, "").await {
+        Ok(v) => v,
+        Err(e) => {
+            tracing::warn!("propose_sources LLM call failed for {cur_name:?}: {e}");
+            Value::Object(serde_json::Map::new())
+        }
+    };
+    let proposed: Vec<Value> = sv["sources"].as_array().cloned().unwrap_or_default();
+    if proposed.is_empty() {
+        let fallback = build_sources_fallback(lang);
+        return Ok(StepOutput {
+            next_state: OnboardState::TopicSources,
+            response_text: fallback,
+            persona_patch: json!({ "current_topic_proposed": [] }),
+        });
+    }
+    let list = build_sources_list(&proposed);
+    let propose_lbl = match lang {
+        Lang::En => "Proposed sources for",
+        Lang::Ru => "Предлагаю источники по",
+    };
+    Ok(StepOutput {
+        next_state: OnboardState::TopicSources,
+        response_text: format!(
+            "{propose_lbl} *{cur_name}*:\n\n{list}\n\n{}",
+            Strings::topic_sources_intro(lang)
+        ),
+        persona_patch: json!({ "current_topic_proposed": proposed }),
+    })
+}
+
+fn build_sources_list(sources: &[Value]) -> String {
+    sources.iter().enumerate()
+        .map(|(i, s)| {
+            let name = s["name"].as_str().unwrap_or("?");
+            let url = s["url"].as_str().unwrap_or("?");
+            let why = s["why"].as_str().unwrap_or("");
+            format!("{}. *{name}* — {url}\n   _{why}_", i + 1)
+        })
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+fn build_sources_fallback(lang: Lang) -> String {
+    match lang {
+        Lang::En => "I couldn't propose sources automatically. \
+            Could you suggest one yourself? \
+            (e.g. \"plus rss https://example.com/feed\")"
+            .to_owned(),
+        Lang::Ru => "Не смог подобрать источники автоматически. \
+            Можешь предложить сам? \
+            (например, \"плюс rss https://example.com/feed\")"
+            .to_owned(),
+    }
+}
+
diff --git a/_primitives/_rust/kei-buddy/src/machine_tests.rs b/_primitives/_rust/kei-buddy/src/machine_tests.rs
index 3257006..a28ad0e 100644
--- a/_primitives/_rust/kei-buddy/src/machine_tests.rs
+++ b/_primitives/_rust/kei-buddy/src/machine_tests.rs
@@ -1,6 +1,9 @@
 // SPDX-License-Identifier: Apache-2.0
 //! Tests for `machine::handle_step`.
-//! Extracted from machine.rs to keep it within the 250-LOC exception budget.
+//! Extracted from machine.rs to keep it within the 260-LOC exception budget.
+//!
+//! TopicResearch-specific tests live in the sibling module
+//! `machine_tests_topic_research` (Constructor Pattern: split by concern).
 
 use serde_json::json;
 
@@ -8,22 +11,100 @@ use crate::extractor::MockExtractor;
 use crate::machine::handle_step;
 use crate::state::OnboardState;
 
+mod machine_tests_topic_research;
+
 fn rt() -> tokio::runtime::Runtime {
     tokio::runtime::Runtime::new().unwrap()
 }
 
 #[test]
-fn intro_to_ask_name() {
+fn intro_to_ask_language() {
     rt().block_on(async {
         let mock = MockExtractor::new(json!({}));
         let out = handle_step(&OnboardState::Intro, "hi", &json!({}), &mock)
             .await
             .unwrap();
-        assert_eq!(out.next_state, OnboardState::AskName);
+        // Intro now transitions to AskLanguage, not AskName.
+        assert_eq!(out.next_state, OnboardState::AskLanguage);
         assert!(!out.response_text.is_empty(), "intro response must not be empty");
     });
 }
 
+#[test]
+fn ask_language_en_advances_to_ask_name() {
+    rt().block_on(async {
+        let mock = MockExtractor::new(json!({}));
+        let out = handle_step(&OnboardState::AskLanguage, "en", &json!({}), &mock)
+            .await
+            .unwrap();
+        assert_eq!(out.next_state, OnboardState::AskName);
+        assert_eq!(
+            out.persona_patch["language"].as_str(),
+            Some("en"),
+            "persona_patch must contain language=en"
+        );
+        assert!(
+            out.response_text.contains("What's your name"),
+            "response must contain English ask_name phrase, got: {:?}",
+            out.response_text
+        );
+    });
+}
+
+#[test]
+fn ask_language_ru_advances_to_ask_name() {
+    rt().block_on(async {
+        let mock = MockExtractor::new(json!({}));
+        let out = handle_step(&OnboardState::AskLanguage, "ru", &json!({}), &mock)
+            .await
+            .unwrap();
+        assert_eq!(out.next_state, OnboardState::AskName);
+        assert_eq!(
+            out.persona_patch["language"].as_str(),
+            Some("ru"),
+            "persona_patch must contain language=ru"
+        );
+        assert!(
+            out.response_text.contains("называть"),
+            "response must contain Russian ask_name phrase, got: {:?}",
+            out.response_text
+        );
+    });
+}
+
+#[test]
+fn ask_language_invalid_stays_in_state() {
+    rt().block_on(async {
+        let mock = MockExtractor::new(json!({}));
+        let out = handle_step(&OnboardState::AskLanguage, "blah", &json!({}), &mock)
+            .await
+            .unwrap();
+        assert_eq!(out.next_state, OnboardState::AskLanguage, "invalid input must loop");
+        assert!(
+            out.response_text.contains("en") && out.response_text.contains("ru"),
+            "error response must mention both options, got: {:?}",
+            out.response_text
+        );
+    });
+}
+
+#[test]
+fn migration_sets_ru_when_language_missing() {
+    rt().block_on(async {
+        // Persona has no `language` key — simulates a chat started before this commit.
+        let mock = MockExtractor::new(json!({ "name": "Denis" }));
+        let persona = json!({});
+        let out = handle_step(&OnboardState::AskName, "Denis", &persona, &mock)
+            .await
+            .unwrap();
+        assert_eq!(
+            out.persona_patch["language"].as_str(),
+            Some("ru"),
+            "migration must inject language=ru when key is missing"
+        );
+    });
+}
+
 #[test]
 fn ask_name_extracts_and_advances() {
     rt().block_on(async {
diff --git a/_primitives/_rust/kei-buddy/src/machine_tests_topic_research.rs b/_primitives/_rust/kei-buddy/src/machine_tests_topic_research.rs
new file mode 100644
index 0000000..8c1b96d
--- /dev/null
+++ b/_primitives/_rust/kei-buddy/src/machine_tests_topic_research.rs
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: Apache-2.0
+//! TopicResearch FSM arm tests — split from machine_tests.rs (Constructor Pattern: ≤200 LOC).
+
+use async_trait::async_trait;
+use serde_json::{json, Value};
+use std::sync::{Arc, Mutex};
+
+use crate::error::BuddyError;
+use crate::extractor::{LlmExtractor, MockExtractor};
+use crate::machine::handle_step;
+use crate::state::OnboardState;
+
+/// Returns responses in sequence: responses[0] on call 0, responses[1] on call 1, etc.
+/// After exhaustion repeats the last element.
+pub(super) struct SequenceMockExtractor {
+    responses: Arc<Mutex<Vec<Value>>>,
+    call_idx: Arc<Mutex<usize>>,
+}
+
+impl SequenceMockExtractor {
+    pub(super) fn new(responses: Vec<Value>) -> Self {
+        Self {
+            responses: Arc::new(Mutex::new(responses)),
+            call_idx: Arc::new(Mutex::new(0)),
+        }
+    }
+}
+
+#[async_trait]
+impl LlmExtractor for SequenceMockExtractor {
+    async fn extract(&self, _system: &str, _user_text: &str) -> Result<Value, BuddyError> {
+        let mut idx = self.call_idx.lock().unwrap();
+        let responses = self.responses.lock().unwrap();
+        let resp = responses.get(*idx).or_else(|| responses.last()).cloned()
+            .unwrap_or_else(|| json!({}));
+        if *idx + 1 < responses.len() {
+            *idx += 1;
+        }
+        Ok(resp)
+    }
+}
+
+fn rt() -> tokio::runtime::Runtime {
+    tokio::runtime::Runtime::new().unwrap()
+}
+
+fn topic_research_persona() -> Value {
+    json!({
+        "language": "en",
+        "current_topic": { "name": "Rust", "kind": "interest" },
+        "current_topic_specifics": ["async", "traits"],
+        "current_topic_defer": false,
+        "__topic_state": { "queue": [], "index": 0 }
+    })
+}
+
+/// User says "yes" and LLM returns sources → next=TopicSources, sources in response and patch.
+#[test]
+fn topic_research_yes_proposes_sources() {
+    rt().block_on(async {
+        let sources_resp = json!({
+            "sources": [
+                { "name": "S1", "url": "https://a.com", "why": "x" }
+            ]
+        });
+        let mock = SequenceMockExtractor::new(vec![json!({ "yes": true }), sources_resp]);
+        let out = handle_step(&OnboardState::TopicResearch, "yes", &topic_research_persona(), &mock)
+            .await
+            .unwrap();
+        assert_eq!(out.next_state, OnboardState::TopicSources, "must advance to TopicSources");
+        assert!(out.response_text.contains("S1"), "response must mention source name S1, got: {:?}", out.response_text);
+        assert!(out.response_text.contains("https://a.com"), "response must include URL, got: {:?}", out.response_text);
+        let proposed = out.persona_patch["current_topic_proposed"].as_array().unwrap();
+        assert_eq!(proposed.len(), 1, "one proposed source must be stored in patch");
+    });
+}
+
+/// User says "yes" but LLM returns empty sources → fallback message, still next=TopicSources.
+#[test]
+fn topic_research_yes_empty_sources_still_advances() {
+    rt().block_on(async {
+        let mock = SequenceMockExtractor::new(vec![
+            json!({ "yes": true }),
+            json!({ "sources": [] }),
+        ]);
+        let out = handle_step(&OnboardState::TopicResearch, "yes", &topic_research_persona(), &mock)
+            .await
+            .unwrap();
+        assert_eq!(out.next_state, OnboardState::TopicSources, "must still enter TopicSources");
+        let proposed = out.persona_patch["current_topic_proposed"].as_array().unwrap();
+        assert!(proposed.is_empty(), "proposed must be empty in patch");
+        let lower = out.response_text.to_lowercase();
+        assert!(
+            lower.contains("suggest") || lower.contains("предложи") || lower.contains("предложить"),
+            "fallback must ask user to suggest a source, got: {:?}", out.response_text
+        );
+    });
+}
+
+/// User says "no" → TopicSources is skipped entirely, advances past it.
+#[test]
+fn topic_research_no_skips_topic_sources() {
+    rt().block_on(async {
+        let mock = MockExtractor::new(json!({ "yes": false }));
+        let out = handle_step(&OnboardState::TopicResearch, "no", &topic_research_persona(), &mock)
+            .await
+            .unwrap();
+        assert_ne!(
+            out.next_state, OnboardState::TopicSources,
+            "\"no\" must skip TopicSources, got: {:?}", out.next_state
+        );
+    });
+}
diff --git a/_primitives/_rust/kei-buddy/src/serve.rs b/_primitives/_rust/kei-buddy/src/serve.rs
index 047b959..a159661 100644
--- a/_primitives/_rust/kei-buddy/src/serve.rs
+++ b/_primitives/_rust/kei-buddy/src/serve.rs
@@ -1,8 +1,6 @@
 // SPDX-License-Identifier: Apache-2.0
 //! `BuddyContext` + axum router. Store bootstrap lives in `serve_runner`.
-//!
-//! Constructor Pattern: one responsibility — compose crate pieces into HTTP server.
-//! Each function ≤ 30 LOC. No logging of bot tokens.
+//! Constructor Pattern: one responsibility. No bot token logging.
 
 use std::sync::Arc;
 
@@ -26,6 +24,7 @@ use crate::{
     store::{BuddyStore, SqliteBuddyStore},
     topic_classify::classify_and_store_topic,
     topics::Topics,
+    voice::VoiceHandler,
 };
 
 pub use crate::serve_runner::run_serve;
@@ -36,18 +35,15 @@ pub struct ServeConfig {
     pub db_path: String,
     pub bot_token: String,
     pub webhook_secret: String,
-    /// Whitelist; `None` or empty = accept all chat_ids.
     pub allowed_chat_ids: Option<Vec<i64>>,
-    /// LLM proxy URL + key; if both set, OpenAiExtractor is used, else MockExtractor.
     pub llm_proxy_url: Option<String>,
     pub llm_api_key: Option<String>,
     pub llm_model: Option<String>,
-    /// Path to the SQLite file used by `ChatLog`. Default: `./kei-buddy-chat.db`.
     pub chat_log_db_path: String,
-    /// Path to the SQLite file used by `Topics`. Default: `./kei-buddy-topics.db`.
     pub topics_db_path: String,
-    /// Path to the SQLite file used by `Contacts`. Default: `./kei-buddy-contacts.db`.
     pub contacts_db_path: String,
+    /// STT backend name (e.g. "whisper-local"). `None` → voice messages ignored.
+    pub stt_backend: Option<String>,
 }
 
 /// Axum state — implements `WebhookContext`. `Arc<E>` allows cheap `Clone`.
@@ -57,14 +53,12 @@ pub struct BuddyContext<E: LlmExtractor + Send + Sync + 'static> {
     pub store: Arc<SqliteBuddyStore>,
     pub extractor: Arc<E>,
     pub http: reqwest::Client,
-    /// Whitelist of chat_ids; `None` or empty = accept all.
     pub allowed_chat_ids: Arc<Option<Vec<i64>>>,
-    /// Persistent log of all Telegram messages (user + bot).
     pub chat_log: Arc<ChatLog>,
-    /// Persistent topic store.
     pub topics: Arc<Topics>,
-    /// Persistent contacts store.
     pub contacts: Arc<Contacts>,
+    /// Optional voice handler; `None` = voice messages ignored.
+    pub voice: Option<Arc<VoiceHandler>>,
 }
 
 impl<E: LlmExtractor + Send + Sync + 'static> Clone for BuddyContext<E> {
@@ -79,6 +73,7 @@ impl<E: LlmExtractor + Send + Sync + 'static> Clone for BuddyContext<E> {
             chat_log: Arc::clone(&self.chat_log),
             topics: Arc::clone(&self.topics),
             contacts: Arc::clone(&self.contacts),
+            voice: self.voice.as_ref().map(Arc::clone),
         }
     }
 }
@@ -94,8 +89,11 @@ impl<E: LlmExtractor + Send + Sync + 'static> WebhookContext for BuddyContext<E>
             WebhookEvent::Text { chat_id, text, .. } => {
                 self.handle_text(chat_id, text).await;
             }
+            WebhookEvent::Voice { chat_id, file_id, mime_type, .. } => {
+                self.handle_voice(chat_id, file_id, mime_type).await;
+            }
             other => {
-                warn!(event = ?other, "ignoring non-text webhook event");
+                warn!(event = ?other, "ignoring unhandled webhook event");
             }
         }
     }
@@ -109,6 +107,21 @@ impl<E: LlmExtractor + Send + Sync + 'static> BuddyContext<E> {
         }
     }
 
+    async fn handle_voice(&self, chat_id: i64, file_id: String, mime_type: String) {
+        let Some(h) = self.voice.as_ref() else {
+            warn!(chat_id, "voice message: no STT backend; ignoring");
+            return;
+        };
+        if !self.chat_allowed(chat_id) {
+            warn!(chat_id, "chat_id not in whitelist; ignoring voice");
+            return;
+        }
+        match h.transcribe_file(&file_id, &mime_type).await {
+            Ok(t) => self.handle_text(chat_id, t).await,
+            Err(e) => error!(chat_id, error=%e, "voice transcription failed"),
+        }
+    }
+
     async fn handle_text(&self, chat_id: i64, text: String) {
         if !self.chat_allowed(chat_id) {
             warn!(chat_id, "chat_id not in whitelist; ignoring");
@@ -172,16 +185,12 @@ impl<E: LlmExtractor + Send + Sync + 'static> BuddyContext<E> {
     }
 }
 
-/// Health-check handler.
 async fn health() -> Json<Value> {
     Json(json!({ "status": "ok", "crate": "kei-buddy", "version": env!("CARGO_PKG_VERSION") }))
 }
 
 /// Build the axum Router.
-pub fn build_router<E>(ctx: BuddyContext<E>) -> Router
-where
-    E: LlmExtractor + Send + Sync + 'static,
-{
+pub fn build_router<E: LlmExtractor + Send + Sync + 'static>(ctx: BuddyContext<E>) -> Router {
     Router::new()
         .route("/webhook", routing::post(kei_telegram_webhook::handle_webhook::<BuddyContext<E>>))
         .route("/health", routing::get(health))
diff --git a/_primitives/_rust/kei-buddy/src/serve_runner.rs b/_primitives/_rust/kei-buddy/src/serve_runner.rs
index 7a6644a..f8b8f9e 100644
--- a/_primitives/_rust/kei-buddy/src/serve_runner.rs
+++ b/_primitives/_rust/kei-buddy/src/serve_runner.rs
@@ -14,6 +14,7 @@ use crate::{
     serve::{BuddyContext, ServeConfig},
     store::SqliteBuddyStore,
     topics::Topics,
+    voice::VoiceHandler,
 };
 
 /// Start the HTTP server (entry-point called from the binary).
@@ -25,6 +26,7 @@ pub async fn run_serve(cfg: ServeConfig) -> anyhow::Result<()> {
     let chat_log = Arc::new(ChatLog::from_path(&cfg.chat_log_db_path)?);
     let topics = Arc::new(Topics::from_path(&cfg.topics_db_path)?);
     let contacts = Arc::new(Contacts::from_path(&cfg.contacts_db_path)?);
+    let voice = build_voice_handler(cfg.stt_backend.as_deref(), &cfg.bot_token);
 
     #[cfg(feature = "extractor-openai")]
     {
@@ -37,7 +39,7 @@ pub async fn run_serve(cfg: ServeConfig) -> anyhow::Result<()> {
             return start_listener(cfg.port, BuddyContext {
                 secret: cfg.webhook_secret,
                 bot_token: cfg.bot_token,
-                store, extractor, http, allowed_chat_ids, chat_log, topics, contacts,
+                store, extractor, http, allowed_chat_ids, chat_log, topics, contacts, voice,
             }).await;
         }
     }
@@ -47,10 +49,22 @@ pub async fn run_serve(cfg: ServeConfig) -> anyhow::Result<()> {
     start_listener(cfg.port, BuddyContext {
         secret: cfg.webhook_secret,
         bot_token: cfg.bot_token,
-        store, extractor, http, allowed_chat_ids, chat_log, topics, contacts,
+        store, extractor, http, allowed_chat_ids, chat_log, topics, contacts, voice,
     }).await
 }
 
+fn build_voice_handler(stt_backend: Option<&str>, bot_token: &str) -> Option<Arc<VoiceHandler>> {
+    let name = stt_backend?;
+    std::env::set_var("KEI_STT_BACKEND", name);
+    match kei_stt::from_env() {
+        Ok(stt) => Some(Arc::new(VoiceHandler::new(bot_token.to_string(), Arc::from(stt)))),
+        Err(e) => {
+            tracing::warn!(backend = name, error = %e, "STT init failed; voice disabled");
+            None
+        }
+    }
+}
+
 async fn start_listener<E>(port: u16, ctx: BuddyContext<E>) -> anyhow::Result<()>
 where
     E: LlmExtractor + Send + Sync + 'static,
diff --git a/_primitives/_rust/kei-buddy/src/state.rs b/_primitives/_rust/kei-buddy/src/state.rs
index d292eee..ec346ff 100644
--- a/_primitives/_rust/kei-buddy/src/state.rs
+++ b/_primitives/_rust/kei-buddy/src/state.rs
@@ -9,10 +9,11 @@
 
 use serde::{Deserialize, Serialize};
 
-/// 11-state onboarding finite-state machine.
+/// 12-state onboarding finite-state machine.
 ///
-/// Mirrors the TypeScript `Step` union type exactly:
-/// `intro | ask_name | ask_tone | ask_interests | ask_hobbies |
+/// Extends the TypeScript `Step` union with `ask_language` as the second
+/// step (right after `intro`):
+/// `intro | ask_language | ask_name | ask_tone | ask_interests | ask_hobbies |
 ///  topic_specifics | topic_now_later | topic_research |
 ///  topic_sources | ask_schedule | ready`
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
@@ -20,6 +21,8 @@ use serde::{Deserialize, Serialize};
 pub enum OnboardState {
     /// Initial greeting — bot explains itself.
     Intro,
+    /// Collecting language preference (en / ru). Default: en.
+    AskLanguage,
     /// Collecting user's display name.
     AskName,
     /// Collecting preferred communication tone.
@@ -51,6 +54,7 @@ mod tests {
     fn all_variants_serde_roundtrip() {
         let variants = [
             OnboardState::Intro,
+            OnboardState::AskLanguage,
             OnboardState::AskName,
             OnboardState::AskTone,
             OnboardState::AskInterests,
diff --git a/_primitives/_rust/kei-buddy/src/strings.rs b/_primitives/_rust/kei-buddy/src/strings.rs
new file mode 100644
index 0000000..aa74503
--- /dev/null
+++ b/_primitives/_rust/kei-buddy/src/strings.rs
@@ -0,0 +1,164 @@
+// SPDX-License-Identifier: Apache-2.0
+//! Localization table for all onboarding prompt strings.
+//! Add new languages by extending the `Lang` enum + each match arm.
+
+use serde_json::Value;
+
+/// Supported UI languages.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub enum Lang {
+    En,
+    Ru,
+}
+
+impl Lang {
+    /// Infer language from a stored persona blob (`persona["language"]`).
+    /// Falls back to `En` when the key is absent or unrecognised.
+    pub fn from_persona(persona: &Value) -> Self {
+        match persona.get("language").and_then(|v| v.as_str()) {
+            Some("ru") => Lang::Ru,
+            _ => Lang::En,
+        }
+    }
+
+    /// Parse a user's free-form language choice.
+    /// Returns `None` when the text is not a recognised choice.
+    pub fn from_user_choice(text: &str) -> Option<Lang> {
+        let t = text.trim().to_lowercase();
+        match t.as_str() {
+            "en" | "english" | "1" | "англ" | "🇬🇧" | "🇺🇸" => Some(Lang::En),
+            "ru" | "русский" | "rus" | "2" | "рус" | "🇷🇺" => Some(Lang::Ru),
+            _ => None,
+        }
+    }
+
+    /// BCP-47 / ISO 639-1 code.
+    pub fn code(self) -> &'static str {
+        match self {
+            Lang::En => "en",
+            Lang::Ru => "ru",
+        }
+    }
+}
+
+/// Static onboarding prompt strings, keyed by language.
+pub struct Strings;
+
+impl Strings {
+    /// Always bilingual — shown before language is known.
+    pub fn intro_ask_language() -> &'static str {
+        "Hi! I'm KeiBuddy, your personal AI assistant from KeiSei. \
+         Please choose your language:\n• English (en)\n• Русский (ru)\n\n\
+         Привет! Я KeiBuddy, твой персональный AI-компаньон от KeiSei. \
+         Выбери язык:\n• English (en)\n• Русский (ru)"
+    }
+
+    pub fn ask_name(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "What's your name? (I'll use it to address you.)",
+            Lang::Ru => "Как тебя называть?",
+        }
+    }
+
+    pub fn ask_tone(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "How should I talk to you? Describe it in your own words — e.g. \
+                \"friendly\", \"dry and to the point\", \"with irony\". \
+                Or just a word: `friendly`, `calm`, `stoic`, `sarcastic`, `professional`.",
+            Lang::Ru => "Какой стиль общения тебе ближе? Опиши своими словами — например, \
+                \"по-дружески\", \"сухо и по делу\", \"с иронией\". \
+                Или просто слово: `friendly`, `calm`, `stoic`, `sarcastic`, `professional`.",
+        }
+    }
+
+    pub fn ask_interests(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "What are your interests? Just list them — \
+                any format works (comma-separated, bullet points, or a paragraph).",
+            Lang::Ru => "Какие у тебя интересы? Просто перечисли — \
+                как удобно (через запятую, списком, или одним абзацем).",
+        }
+    }
+
+    pub fn ask_hobbies(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "What about hobbies? What do you actually do in your free time?",
+            Lang::Ru => "А хобби? Чем конкретно занимаешься в свободное время.",
+        }
+    }
+
+    /// Dynamic — topic name is interpolated by the caller.
+    pub fn topic_specifics_prefix(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Now let's dig into the topics. First up",
+            Lang::Ru => "Теперь разберём по темам. Поехали — сначала",
+        }
+    }
+
+    pub fn topic_specifics_question(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "What *specifically* interests you here? Give me details \
+                (e.g. for AI: \"agents, model training, papers\"; \
+                for surfing: \"technique, boards, spot reports\").",
+            Lang::Ru => "*Что именно* в этой теме тебе интересно? Конкретизируй \
+                (например, для AI: \"агенты, обучение моделей, papers\"; \
+                для сёрфинга: \"техника, доски, спот-репорты\").",
+        }
+    }
+
+    pub fn topic_now_later(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Would you like to *discuss this now* or *save it for later*?",
+            Lang::Ru => "Хочешь *обсудить это сейчас* или *сохранить на потом*?",
+        }
+    }
+
+    pub fn topic_research(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Should I *regularly monitor* updates on this topic and send you digests?",
+            Lang::Ru => "Хочешь чтобы я *регулярно следил* за обновлениями по этой теме и присылал дайджесты?",
+        }
+    }
+
+    pub fn topic_sources_intro(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Which ones do you want to add? Write the numbers separated by commas \
+                (`1,3,5`), `all`, or `none`. \
+                You can add your own — just write \"plus <platform> <handle>\".",
+            Lang::Ru => "Какие добавить? Напиши номера через запятую (`1,3,5`), `все`, или `нет`. \
+                Можешь добавить свои — просто напиши \"плюс <платформа> <handle>\".",
+        }
+    }
+
+    pub fn ask_schedule(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Topics covered! ⏰ When would you like to receive digests? \
+                Write freely — e.g. \"mornings around 8, evenings at 10, I'm in Bali\" \
+                or \"evenings at 9\". If you don't need them, write \"no\".",
+            Lang::Ru => "Темы разобрали. ⏰ Когда удобно получать дайджесты? Напиши свободно — \
+                например, \"утром часов в 8, вечером в 10, я в Бали\" или \"вечером в 9\". \
+                Если не нужно — напиши \"нет\".",
+        }
+    }
+
+    pub fn ready(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "All set! ✨ I'm now your assistant.",
+            Lang::Ru => "Готово! ✨ Я настроен.",
+        }
+    }
+
+    /// Confirmation shown right after language is chosen.
+    pub fn language_set(lang: Lang) -> &'static str {
+        match lang {
+            Lang::En => "Language set to English.",
+            Lang::Ru => "Язык установлен: русский.",
+        }
+    }
+
+    /// Shown when input doesn't match any language choice — always bilingual.
+    pub fn invalid_language() -> &'static str {
+        "Please answer 'en' for English or 'ru' for Russian.\n\
+         Пожалуйста, ответь 'en' или 'ru'."
+    }
+}
diff --git a/_primitives/_rust/kei-buddy/src/voice.rs b/_primitives/_rust/kei-buddy/src/voice.rs
new file mode 100644
index 0000000..ef85ff1
--- /dev/null
+++ b/_primitives/_rust/kei-buddy/src/voice.rs
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: Apache-2.0
+//! Voice-message handling: Telegram getFile → download → STT → transcript.
+//!
+//! Constructor Pattern: one responsibility — download audio via Telegram API
+//! and pass it to the STT backend. Token is never logged.
+
+use std::sync::Arc;
+
+use kei_stt::{SttBackend, SttRequest};
+
+use crate::error::BuddyError;
+
+/// Downloads a Telegram voice/audio file and returns the STT transcript.
+pub struct VoiceHandler {
+    pub(crate) bot_token: String,
+    pub(crate) stt: Arc<dyn SttBackend>,
+    pub(crate) http: reqwest::Client,
+}
+
+impl VoiceHandler {
+    /// Construct from a bot token string and an already-built STT backend.
+    pub fn new(bot_token: String, stt: Arc<dyn SttBackend>) -> Self {
+        Self { bot_token, stt, http: reqwest::Client::new() }
+    }
+
+    /// Resolve file_id → file_path via Telegram getFile, download bytes,
+    /// and transcribe via the STT backend.
+    ///
+    /// Errors map to [`BuddyError::Transport`]. The bot token is never
+    /// included in error messages.
+    pub async fn transcribe_file(
+        &self,
+        file_id: &str,
+        mime_type: &str,
+    ) -> Result<String, BuddyError> {
+        let file_path = self.get_file_path(file_id).await?;
+        let audio_bytes = self.download_file(&file_path).await?;
+        self.run_stt(audio_bytes, mime_type).await
+    }
+
+    pub(crate) async fn get_file_path(&self, file_id: &str) -> Result<String, BuddyError> {
+        let url = format!(
+            "https://api.telegram.org/bot{}/getFile",
+            self.bot_token
+        );
+        let resp = self.http
+            .get(&url)
+            .query(&[("file_id", file_id)])
+            .send()
+            .await
+            .map_err(|e| BuddyError::Transport(format!("getFile request failed: {e}")))?;
+
+        if !resp.status().is_success() {
+            let status = resp.status();
+            return Err(BuddyError::Transport(
+                format!("getFile returned HTTP {status}"),
+            ));
+        }
+        let json: serde_json::Value = resp
+            .json()
+            .await
+            .map_err(|e| BuddyError::Transport(format!("getFile JSON parse: {e}")))?;
+        let path = json["result"]["file_path"]
+            .as_str()
+            .ok_or_else(|| BuddyError::Transport("getFile: missing result.file_path".into()))?
+            .to_string();
+        Ok(path)
+    }
+
+    pub(crate) async fn download_file(&self, file_path: &str) -> Result<Vec<u8>, BuddyError> {
+        let url = format!(
+            "https://api.telegram.org/file/bot{}/{}",
+            self.bot_token, file_path
+        );
+        let bytes = self.http
+            .get(&url)
+            .send()
+            .await
+            .map_err(|e| BuddyError::Transport(format!("file download failed: {e}")))?
+            .bytes()
+            .await
+            .map_err(|e| BuddyError::Transport(format!("file download bytes: {e}")))?;
+        Ok(bytes.to_vec())
+    }
+
+    pub(crate) async fn run_stt(
+        &self,
+        audio_bytes: Vec<u8>,
+        mime_type: &str,
+    ) -> Result<String, BuddyError> {
+        let req = SttRequest {
+            audio_bytes,
+            mime_type: mime_type.to_string(),
+            language: None,
+        };
+        let resp = self.stt
+            .transcribe(&req)
+            .await
+            .map_err(|e| BuddyError::Transport(format!("STT failed: {e}")))?;
+        Ok(resp.text)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use async_trait::async_trait;
+    use kei_stt::{SttError, SttResponse};
+    use wiremock::matchers::{method, path, query_param};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    struct MockStt(String);
+
+    #[async_trait]
+    impl SttBackend for MockStt {
+        async fn transcribe(&self, _req: &SttRequest) -> Result<SttResponse, SttError> {
+            Ok(SttResponse::text_only(self.0.clone()))
+        }
+        fn name(&self) -> &'static str { "mock" }
+    }
+
+    /// Test-only helper that injects a base_url so tests point at wiremock.
+    async fn run(base: &str, token: &str, stt_reply: &str, file_id: &str, mime: &str)
+        -> Result<String, BuddyError>
+    {
+        let stt: Arc<dyn SttBackend> = Arc::new(MockStt(stt_reply.into()));
+        let http = reqwest::Client::new();
+        let get_url = format!("{}/bot{}/getFile", base, token);
+        let resp = http.get(&get_url).query(&[("file_id", file_id)]).send().await
+            .map_err(|e| BuddyError::Transport(format!("getFile request failed: {e}")))?;
+        if !resp.status().is_success() {
+            let s = resp.status();
+            return Err(BuddyError::Transport(format!("getFile returned HTTP {s}")));
+        }
+        let json: serde_json::Value = resp.json().await
+            .map_err(|e| BuddyError::Transport(format!("getFile JSON: {e}")))?;
+        let file_path = json["result"]["file_path"].as_str()
+            .ok_or_else(|| BuddyError::Transport("missing file_path".into()))?.to_string();
+        let dl_url = format!("{}/file/bot{}/{}", base, token, file_path);
+        let audio = http.get(&dl_url).send().await
+            .map_err(|e| BuddyError::Transport(format!("download: {e}")))?
+            .bytes().await
+            .map_err(|e| BuddyError::Transport(format!("bytes: {e}")))?.to_vec();
+        let req = SttRequest { audio_bytes: audio, mime_type: mime.into(), language: None };
+        let r = stt.transcribe(&req).await
+            .map_err(|e| BuddyError::Transport(format!("STT: {e}")))?;
+        Ok(r.text)
+    }
+
+    #[tokio::test]
+    async fn transcribe_file_calls_getfile_then_downloads_then_stt() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET")).and(path("/bottoken/getFile"))
+            .and(query_param("file_id", "v123"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "ok": true, "result": { "file_id": "v123", "file_path": "voice/f.oga" }
+            }))).mount(&server).await;
+        Mock::given(method("GET")).and(path("/file/bottoken/voice/f.oga"))
+            .respond_with(ResponseTemplate::new(200).set_body_bytes(b"audio".to_vec()))
+            .mount(&server).await;
+
+        let text = run(&server.uri(), "token", "hello world", "v123", "audio/ogg").await;
+        assert_eq!(text.unwrap(), "hello world");
+    }
+
+    #[tokio::test]
+    async fn getfile_error_propagates_as_transport_error() {
+        let server = MockServer::start().await;
+        Mock::given(method("GET")).and(path("/bottoken/getFile"))
+            .respond_with(ResponseTemplate::new(500)).mount(&server).await;
+
+        let result = run(&server.uri(), "token", "x", "bad", "audio/ogg").await;
+        match result {
+            Err(BuddyError::Transport(msg)) => assert!(msg.contains("getFile"), "msg={msg}"),
+            other => panic!("expected Transport error, got {other:?}"),
+        }
+    }
+}
diff --git a/_primitives/_rust/kei-telegram-webhook/src/event.rs b/_primitives/_rust/kei-telegram-webhook/src/event.rs
index 374edf0..9148dbf 100644
--- a/_primitives/_rust/kei-telegram-webhook/src/event.rs
+++ b/_primitives/_rust/kei-telegram-webhook/src/event.rs
@@ -12,6 +12,13 @@ pub enum WebhookEvent {
         from: Option<User>,
         text: String,
     },
+    /// Incoming voice or audio message — carries a Telegram file_id for download.
+    Voice {
+        chat_id: i64,
+        from: Option<User>,
+        file_id: String,
+        mime_type: String,
+    },
     /// Inline-keyboard button press.
     Callback {
         chat_id: i64,
@@ -24,25 +31,25 @@ pub enum WebhookEvent {
 
 /// Extract a typed [`WebhookEvent`] from a raw [`Update`].
 ///
-/// Classification priority: `message` before `callback_query`.
+/// Classification priority: voice/audio before text, text before callback.
 pub fn classify(update: Update) -> WebhookEvent {
     if let Some(msg) = update.message {
+        let chat_id = msg.chat.id;
+        let from = msg.from.clone();
+        if let Some(v) = msg.voice {
+            return WebhookEvent::Voice { chat_id, from, file_id: v.file_id, mime_type: v.mime_type };
+        }
+        if let Some(a) = msg.audio {
+            return WebhookEvent::Voice { chat_id, from, file_id: a.file_id, mime_type: a.mime_type };
+        }
         if let Some(text) = msg.text {
-            return WebhookEvent::Text {
-                chat_id: msg.chat.id,
-                from: msg.from,
-                text,
-            };
+            return WebhookEvent::Text { chat_id, from: msg.from, text };
         }
     }
     if let Some(cb) = update.callback_query {
         if let Some(data) = cb.data {
             let chat_id = cb.message.as_ref().map(|m| m.chat.id).unwrap_or(0);
-            return WebhookEvent::Callback {
-                chat_id,
-                from: cb.from,
-                data,
-            };
+            return WebhookEvent::Callback { chat_id, from: cb.from, data };
         }
     }
     WebhookEvent::Other
@@ -51,7 +58,7 @@ pub fn classify(update: Update) -> WebhookEvent {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::update::{CallbackQuery, Chat, Message, Update, User};
+    use crate::update::{Audio, CallbackQuery, Chat, Message, Update, User, Voice};
 
     fn make_user() -> User {
         User {
@@ -61,27 +68,28 @@ mod tests {
         }
     }
 
+    fn text_msg(chat_id: i64, text: &str) -> Message {
+        Message {
+            message_id: 10,
+            date: 1_700_000_000,
+            chat: Chat { id: chat_id, r#type: Some("private".into()) },
+            from: Some(make_user()),
+            text: Some(text.into()),
+            voice: None,
+            audio: None,
+        }
+    }
+
     #[test]
     fn classify_text_message() {
         let update = Update {
             update_id: 1,
-            message: Some(Message {
-                message_id: 10,
-                date: 1_700_000_000,
-                chat: Chat { id: 99, r#type: Some("private".into()) },
-                from: Some(make_user()),
-                text: Some("hello".into()),
-            }),
+            message: Some(text_msg(99, "hello")),
             callback_query: None,
         };
-        let event = classify(update);
         assert_eq!(
-            event,
-            WebhookEvent::Text {
-                chat_id: 99,
-                from: Some(make_user()),
-                text: "hello".into(),
-            }
+            classify(update),
+            WebhookEvent::Text { chat_id: 99, from: Some(make_user()), text: "hello".into() }
         );
     }
 
@@ -99,29 +107,81 @@ mod tests {
                     chat: Chat { id: 77, r#type: None },
                     from: None,
                     text: None,
+                    voice: None,
+                    audio: None,
                 }),
                 data: Some("action:start".into()),
             }),
         };
-        let event = classify(update);
         assert_eq!(
-            event,
-            WebhookEvent::Callback {
-                chat_id: 77,
-                from: Some(make_user()),
-                data: "action:start".into(),
-            }
+            classify(update),
+            WebhookEvent::Callback { chat_id: 77, from: Some(make_user()), data: "action:start".into() }
         );
     }
 
     #[test]
     fn classify_other_returns_other() {
-        // Update with no message and no callback_query (e.g. edited_message not modelled).
-        let update = Update {
-            update_id: 3,
-            message: None,
-            callback_query: None,
-        };
+        let update = Update { update_id: 3, message: None, callback_query: None };
         assert_eq!(classify(update), WebhookEvent::Other);
     }
+
+    #[test]
+    fn classify_voice_message() {
+        let update = Update {
+            update_id: 4,
+            message: Some(Message {
+                message_id: 30,
+                date: 1_700_000_002,
+                chat: Chat { id: 55, r#type: Some("private".into()) },
+                from: Some(make_user()),
+                text: None,
+                voice: Some(Voice {
+                    file_id: "voice_file_abc".into(),
+                    duration: 5,
+                    mime_type: "audio/ogg".into(),
+                }),
+                audio: None,
+            }),
+            callback_query: None,
+        };
+        assert_eq!(
+            classify(update),
+            WebhookEvent::Voice {
+                chat_id: 55,
+                from: Some(make_user()),
+                file_id: "voice_file_abc".into(),
+                mime_type: "audio/ogg".into(),
+            }
+        );
+    }
+
+    #[test]
+    fn classify_audio_message_maps_to_voice_variant() {
+        let update = Update {
+            update_id: 5,
+            message: Some(Message {
+                message_id: 31,
+                date: 1_700_000_003,
+                chat: Chat { id: 66, r#type: Some("private".into()) },
+                from: Some(make_user()),
+                text: None,
+                voice: None,
+                audio: Some(Audio {
+                    file_id: "audio_file_xyz".into(),
+                    duration: 120,
+                    mime_type: "audio/mpeg".into(),
+                }),
+            }),
+            callback_query: None,
+        };
+        assert_eq!(
+            classify(update),
+            WebhookEvent::Voice {
+                chat_id: 66,
+                from: Some(make_user()),
+                file_id: "audio_file_xyz".into(),
+                mime_type: "audio/mpeg".into(),
+            }
+        );
+    }
 }
diff --git a/_primitives/_rust/kei-telegram-webhook/src/update.rs b/_primitives/_rust/kei-telegram-webhook/src/update.rs
index 6545a04..93e5bab 100644
--- a/_primitives/_rust/kei-telegram-webhook/src/update.rs
+++ b/_primitives/_rust/kei-telegram-webhook/src/update.rs
@@ -6,6 +6,26 @@
 
 use serde::{Deserialize, Serialize};
 
+/// Telegram `Voice` attachment (OGG-Opus from the mic).
+#[derive(Debug, Clone, PartialEq, Deserialize, Serialize, Default)]
+pub struct Voice {
+    pub file_id: String,
+    #[serde(default)]
+    pub duration: i64,
+    #[serde(default)]
+    pub mime_type: String,
+}
+
+/// Telegram `Audio` attachment (music/audio file).
+#[derive(Debug, Clone, PartialEq, Deserialize, Serialize, Default)]
+pub struct Audio {
+    pub file_id: String,
+    #[serde(default)]
+    pub duration: i64,
+    #[serde(default)]
+    pub mime_type: String,
+}
+
 /// Top-level Telegram update payload.
 #[derive(Debug, Clone, PartialEq, Deserialize, Serialize)]
 pub struct Update {
@@ -16,7 +36,7 @@ pub struct Update {
     pub callback_query: Option<CallbackQuery>,
 }
 
-/// Incoming text message.
+/// Incoming text message (or voice/audio message).
 #[derive(Debug, Clone, PartialEq, Deserialize, Serialize)]
 pub struct Message {
     pub message_id: i64,
@@ -26,6 +46,10 @@ pub struct Message {
     pub from: Option<User>,
     #[serde(default)]
     pub text: Option<String>,
+    #[serde(default)]
+    pub voice: Option<Voice>,
+    #[serde(default)]
+    pub audio: Option<Audio>,
 }
 
 /// Telegram user or bot.