feat(voice): switch voice ASR to Yandex SpeechKit

.env.example

@@ -52,13 +52,13 @@ AI_VOICE_RUNTIME_SERVICE_URL=http://localhost:8018
 AI_VOICE_ENABLED=0
 AI_VOICE_POLICY_MODE=v2_fast_conversational
 AI_VOICE_QUEUE_CONFIG_JSON={}
-AI_VOICE_ASR_PROVIDER=openai
+AI_VOICE_ASR_PROVIDER=yandex
 AI_VOICE_TTS_PROVIDER=yandex
 AI_VOICE_V2_ENABLED=1
 AI_VOICE_V2_QUEUE_CODES=voice_lab_ai
 AI_VOICE_V2_ACK_MODE=immediate_short
 AI_VOICE_V2_DUPLEX_ENABLED=1
-AI_VOICE_V2_STREAMING_ASR_BACKEND=local_sidecar
+AI_VOICE_V2_STREAMING_ASR_BACKEND=yandex_speechkit
 AI_VOICE_V2_STREAMING_ASR_BASE_URL=http://127.0.0.1:8021
 AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS=4
 AI_VOICE_V2_STREAMING_ASR_MODEL=base
@@ -76,6 +76,16 @@ AI_VOICE_AUDIOSOCKET_ENABLED=0
 AI_VOICE_AUDIOSOCKET_HOST=0.0.0.0
 AI_VOICE_AUDIOSOCKET_PORT=9019
 AI_VOICE_ASR_MODEL=gpt-4o-mini-transcribe
+AI_VOICE_ASR_YANDEX_API_BASE=https://stt.api.ml.yandexcloud.kz
+AI_VOICE_ASR_YANDEX_API_VERSION=auto
+AI_VOICE_ASR_YANDEX_API_KEY=
+AI_VOICE_ASR_YANDEX_IAM_TOKEN=
+AI_VOICE_ASR_YANDEX_FOLDER_ID=
+AI_VOICE_ASR_YANDEX_LANGUAGE=ru-RU
+AI_VOICE_ASR_YANDEX_TOPIC=general
+AI_VOICE_ASR_YANDEX_SAMPLE_RATE_HZ=8000
+AI_VOICE_ASR_YANDEX_TIMEOUT_SECONDS=8
+AI_VOICE_ASR_YANDEX_POLL_INTERVAL_SECONDS=0.25
 AI_VOICE_TTS_MODEL=gpt-4o-mini-tts
 AI_VOICE_TTS_VOICE=alloy
 AI_VOICE_TTS_YANDEX_API_BASE=https://tts.api.ml.yandexcloud.kz

deployment/docker-compose.parallel.server.yml

@@ -22,6 +22,9 @@ x-app-env: &app_env
   KB_SERVICE_URL: http://kb-service:8000
   REPORTING_SERVICE_URL: http://reporting-service:8000
   SUPERVISOR_SERVICE_URL: http://supervisor-service:8000
+  AI_VOICE_ASR_PROVIDER: yandex
+  AI_VOICE_TTS_PROVIDER: yandex
+  AI_VOICE_V2_STREAMING_ASR_BACKEND: yandex_speechkit
   AI_VOICE_V2_STREAMING_ASR_BASE_URL: http://streaming-asr-sidecar:8021
   AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS: "4"
 

deployment/docker-compose.server.registry.yml

@@ -22,6 +22,9 @@ x-app-env: &app_env
   KB_SERVICE_URL: http://kb-service:8000
   REPORTING_SERVICE_URL: http://reporting-service:8000
   SUPERVISOR_SERVICE_URL: http://supervisor-service:8000
+  AI_VOICE_ASR_PROVIDER: yandex
+  AI_VOICE_TTS_PROVIDER: yandex
+  AI_VOICE_V2_STREAMING_ASR_BACKEND: yandex_speechkit
   AI_VOICE_V2_STREAMING_ASR_BASE_URL: http://streaming-asr-sidecar:8021
   AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS: "4"
 

deployment/docker-compose.server.yml

@@ -22,6 +22,9 @@ x-app-env: &app_env
   KB_SERVICE_URL: http://kb-service:8000
   REPORTING_SERVICE_URL: http://reporting-service:8000
   SUPERVISOR_SERVICE_URL: http://supervisor-service:8000
+  AI_VOICE_ASR_PROVIDER: yandex
+  AI_VOICE_TTS_PROVIDER: yandex
+  AI_VOICE_V2_STREAMING_ASR_BACKEND: yandex_speechkit
   AI_VOICE_V2_STREAMING_ASR_BASE_URL: http://streaming-asr-sidecar:8021
   AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS: "4"
 

deployment/docker-compose.yml

@@ -19,7 +19,7 @@ x-app-env: &app_env
   AI_VOICE_ENABLED: "0"
   AI_VOICE_QUEUE_CONFIG_JSON: "{}"
   AI_VOICE_POLICY_MODE: v2_fast_conversational
-  AI_VOICE_ASR_PROVIDER: openai
+  AI_VOICE_ASR_PROVIDER: yandex
   AI_VOICE_TTS_PROVIDER: yandex
   AI_VOICE_AUDIOSOCKET_ENABLED: "1"
   AI_VOICE_AUDIOSOCKET_HOST: 0.0.0.0
@@ -35,7 +35,7 @@ x-app-env: &app_env
   AI_VOICE_V2_QUEUE_CODES: voice_lab_ai
   AI_VOICE_V2_ACK_MODE: immediate_short
   AI_VOICE_V2_DUPLEX_ENABLED: "1"
-  AI_VOICE_V2_STREAMING_ASR_BACKEND: local_sidecar
+  AI_VOICE_V2_STREAMING_ASR_BACKEND: yandex_speechkit
   AI_VOICE_V2_STREAMING_ASR_BASE_URL: http://streaming-asr-sidecar:8021
   AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS: "4"
   AI_VOICE_V2_PREBAKED_ACK_ENABLED: "1"

deployment/helm/values.yaml

@@ -47,8 +47,8 @@ recording:
 voiceAI:
   enabled: "0"
   queueConfigJson: "{}"
-  asrProvider: openai
-  ttsProvider: openai
+  asrProvider: yandex
+  ttsProvider: yandex
   audioSocketEnabled: "1"
   audioSocketHost: 0.0.0.0
   audioSocketPort: "9019"

scripts/local_stack.py

@@ -309,14 +309,14 @@ def spawn_service(spec: dict[str, Any], runtime_dir: Path, data_dir: Path, base_
     env["AI_VOICE_RUNTIME_SERVICE_URL"] = env.get("AI_VOICE_RUNTIME_SERVICE_URL", "http://127.0.0.1:8018")
     env["AI_VOICE_ENABLED"] = env.get("AI_VOICE_ENABLED", "0")
     env["AI_VOICE_QUEUE_CONFIG_JSON"] = env.get("AI_VOICE_QUEUE_CONFIG_JSON", "{}")
-    env["AI_VOICE_ASR_PROVIDER"] = env.get("AI_VOICE_ASR_PROVIDER", "openai")
+    env["AI_VOICE_ASR_PROVIDER"] = env.get("AI_VOICE_ASR_PROVIDER", "yandex")
     env["AI_VOICE_TTS_PROVIDER"] = env.get("AI_VOICE_TTS_PROVIDER", "yandex")
     env["AI_VOICE_POLICY_MODE"] = env.get("AI_VOICE_POLICY_MODE", "v2_fast_conversational")
     env["AI_VOICE_V2_ENABLED"] = env.get("AI_VOICE_V2_ENABLED", "1")
     env["AI_VOICE_V2_QUEUE_CODES"] = env.get("AI_VOICE_V2_QUEUE_CODES", "voice_lab_ai")
     env["AI_VOICE_V2_ACK_MODE"] = env.get("AI_VOICE_V2_ACK_MODE", "immediate_short")
     env["AI_VOICE_V2_DUPLEX_ENABLED"] = env.get("AI_VOICE_V2_DUPLEX_ENABLED", "1")
-    env["AI_VOICE_V2_STREAMING_ASR_BACKEND"] = env.get("AI_VOICE_V2_STREAMING_ASR_BACKEND", "local_sidecar")
+    env["AI_VOICE_V2_STREAMING_ASR_BACKEND"] = env.get("AI_VOICE_V2_STREAMING_ASR_BACKEND", "yandex_speechkit")
     env["AI_VOICE_V2_STREAMING_ASR_BASE_URL"] = env.get("AI_VOICE_V2_STREAMING_ASR_BASE_URL", "http://127.0.0.1:8021")
     env["AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS"] = env.get("AI_VOICE_V2_STREAMING_ASR_TIMEOUT_SECONDS", "4")
     env["AI_VOICE_V2_STREAMING_ASR_MODEL"] = env.get("AI_VOICE_V2_STREAMING_ASR_MODEL", "base")

services/ai_voice_runtime_service/app.py

@@ -84,7 +84,7 @@ def _interaction_service_url() -> str:
 
 
 def _asr_provider_name() -> str:
-    return os.getenv("AI_VOICE_ASR_PROVIDER", "openai").strip() or "openai"
+    return os.getenv("AI_VOICE_ASR_PROVIDER", "yandex").strip() or "yandex"
 
 
 def _tts_provider_name() -> str:
@@ -164,7 +164,7 @@ def _voice_v2_duplex_enabled() -> bool:
 
 
 def _voice_v2_streaming_asr_backend() -> str:
-    return str(os.getenv("AI_VOICE_V2_STREAMING_ASR_BACKEND", "local_sidecar") or "local_sidecar").strip().lower()
+    return str(os.getenv("AI_VOICE_V2_STREAMING_ASR_BACKEND", "yandex_speechkit") or "yandex_speechkit").strip().lower()
 
 
 def _voice_v2_prebaked_ack_enabled() -> bool:

services/ai_voice_runtime_service/providers/asr.py

@@ -1,7 +1,12 @@
 from __future__ import annotations
 
 import base64
+import io
 import os
+import threading
+import time
+import uuid
+import wave
 from dataclasses import dataclass
 
 import httpx
@@ -30,6 +35,99 @@ def _openai_asr_model() -> str:
     return os.getenv("AI_VOICE_ASR_MODEL", "gpt-4o-mini-transcribe").strip() or "gpt-4o-mini-transcribe"
 
 
+def _yandex_asr_api_base() -> str:
+    explicit = os.getenv("AI_VOICE_ASR_YANDEX_API_BASE", "").strip()
+    if explicit:
+        return explicit.rstrip("/")
+    tts_base = os.getenv("AI_VOICE_TTS_YANDEX_API_BASE", "https://tts.api.ml.yandexcloud.kz").strip().lower()
+    if "yandexcloud.kz" in tts_base:
+        return "https://stt.api.ml.yandexcloud.kz"
+    return "https://stt.api.cloud.yandex.net"
+
+
+def _yandex_asr_api_key() -> str:
+    return (
+        os.getenv("AI_VOICE_ASR_YANDEX_API_KEY", "").strip()
+        or os.getenv("AI_VOICE_TTS_YANDEX_API_KEY", "").strip()
+    )
+
+
+def _yandex_asr_iam_token() -> str:
+    return (
+        os.getenv("AI_VOICE_ASR_YANDEX_IAM_TOKEN", "").strip()
+        or os.getenv("AI_VOICE_TTS_YANDEX_IAM_TOKEN", "").strip()
+    )
+
+
+def _yandex_asr_folder_id() -> str:
+    return (
+        os.getenv("AI_VOICE_ASR_YANDEX_FOLDER_ID", "").strip()
+        or os.getenv("AI_VOICE_TTS_YANDEX_FOLDER_ID", "").strip()
+    )
+
+
+def _yandex_asr_timeout_seconds() -> float:
+    raw = os.getenv("AI_VOICE_ASR_YANDEX_TIMEOUT_SECONDS", "").strip()
+    if not raw:
+        return _timeout_seconds()
+    try:
+        value = float(raw)
+    except ValueError:
+        value = _timeout_seconds()
+    return max(value, 3.0)
+
+
+def _yandex_asr_sample_rate_hz() -> int:
+    raw = os.getenv("AI_VOICE_ASR_YANDEX_SAMPLE_RATE_HZ", "8000").strip()
+    try:
+        value = int(raw)
+    except ValueError:
+        value = 8000
+    if value < 8000 or value > 48000:
+        return 8000
+    return value
+
+
+def _yandex_asr_topic() -> str:
+    return os.getenv("AI_VOICE_ASR_YANDEX_TOPIC", "general").strip() or "general"
+
+
+def _yandex_asr_api_version(api_base: str) -> str:
+    raw = os.getenv("AI_VOICE_ASR_YANDEX_API_VERSION", "auto").strip().lower() or "auto"
+    if raw in {"v1", "v1_sync", "sync"}:
+        return "v1"
+    if raw in {"v3", "v3_async", "async"}:
+        return "v3"
+    if "yandexcloud.kz" in str(api_base or "").lower():
+        return "v3"
+    return "v1"
+
+
+def _yandex_asr_poll_interval_seconds() -> float:
+    raw = os.getenv("AI_VOICE_ASR_YANDEX_POLL_INTERVAL_SECONDS", "0.25").strip()
+    try:
+        value = float(raw)
+    except ValueError:
+        value = 0.25
+    return max(min(value, 2.0), 0.05)
+
+
+def _yandex_asr_default_language() -> str:
+    return os.getenv("AI_VOICE_ASR_YANDEX_LANGUAGE", "ru-RU").strip() or "ru-RU"
+
+
+def _normalize_yandex_asr_language(language: str | None) -> str:
+    raw = str(language or "").strip()
+    if not raw:
+        return _yandex_asr_default_language()
+    lowered = raw.replace("_", "-").lower()
+    if lowered in {"ru", "ru-ru"}:
+        return "ru-RU"
+    if lowered in {"kk", "kz", "kk-kz", "kz-kz"}:
+        return "kk-KZ"
+    return raw
+
+
 def _streaming_asr_api_base() -> str:
     return (
         os.getenv("AI_VOICE_V2_STREAMING_ASR_BASE_URL", "http://127.0.0.1:8021").strip()
@@ -152,6 +250,214 @@ class OpenAIASRProvider(ASRProvider):
         return self.transcribe(audio_bytes, language_hint=language_hint)
 
 
+class YandexSpeechKitASRProvider(ASRProvider):
+    name = "yandex"
+
+    def __init__(
+        self,
+        *,
+        api_base: str | None = None,
+        api_key: str | None = None,
+        iam_token: str | None = None,
+        folder_id: str | None = None,
+        timeout_seconds: float | None = None,
+        sample_rate_hz: int | None = None,
+        topic: str | None = None,
+    ) -> None:
+        self._api_base = str(api_base or _yandex_asr_api_base()).strip().rstrip("/")
+        self._api_key = str(api_key if api_key is not None else _yandex_asr_api_key()).strip()
+        self._iam_token = str(iam_token if iam_token is not None else _yandex_asr_iam_token()).strip()
+        self._folder_id = str(folder_id if folder_id is not None else _yandex_asr_folder_id()).strip()
+        self._timeout_seconds = max(
+            float(timeout_seconds if timeout_seconds is not None else _yandex_asr_timeout_seconds()),
+            3.0,
+        )
+        self._sample_rate_hz = int(sample_rate_hz if sample_rate_hz is not None else _yandex_asr_sample_rate_hz())
+        self._topic = str(topic if topic is not None else _yandex_asr_topic()).strip()
+        self._api_version = _yandex_asr_api_version(self._api_base)
+        self._poll_interval_seconds = _yandex_asr_poll_interval_seconds()
+
+    @staticmethod
+    def _lpcm_from_audio_bytes(audio_bytes: bytes, *, default_sample_rate_hz: int) -> tuple[bytes, int]:
+        if not audio_bytes:
+            return b"", default_sample_rate_hz
+        try:
+            with wave.open(io.BytesIO(audio_bytes), "rb") as wav_file:
+                sample_width = wav_file.getsampwidth()
+                channels = wav_file.getnchannels()
+                sample_rate = wav_file.getframerate()
+                if sample_width != 2 or channels != 1:
+                    return audio_bytes, default_sample_rate_hz
+                return wav_file.readframes(wav_file.getnframes()), int(sample_rate or default_sample_rate_hz)
+        except (wave.Error, EOFError):
+            return audio_bytes, default_sample_rate_hz
+
+    def _headers(self, *, content_type: str = "application/octet-stream") -> dict[str, str]:
+        if not self._api_key and not self._iam_token:
+            raise RuntimeError(
+                "AI_VOICE_ASR_YANDEX_API_KEY or AI_VOICE_ASR_YANDEX_IAM_TOKEN is required for Yandex ASR"
+            )
+        headers = {"Content-Type": content_type}
+        if self._api_key:
+            headers["Authorization"] = f"Api-Key {self._api_key}"
+        else:
+            headers["Authorization"] = f"Bearer {self._iam_token}"
+        if self._folder_id and self._api_version == "v3":
+            headers["x-folder-id"] = self._folder_id
+        return headers
+
+    def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+        pcm_bytes, sample_rate_hz = self._lpcm_from_audio_bytes(
+            audio_bytes,
+            default_sample_rate_hz=self._sample_rate_hz,
+        )
+        if not pcm_bytes:
+            return ASRTranscription(text="", language=language_hint, confidence=None)
+
+        language = _normalize_yandex_asr_language(language_hint)
+        if self._api_version == "v3":
+            return self._transcribe_v3_async(pcm_bytes, sample_rate_hz=sample_rate_hz, language=language)
+        return self._transcribe_v1_sync(pcm_bytes, sample_rate_hz=sample_rate_hz, language=language)
+
+    def _transcribe_v1_sync(self, pcm_bytes: bytes, *, sample_rate_hz: int, language: str) -> ASRTranscription:
+        params: dict[str, str] = {
+            "lang": language,
+            "format": "lpcm",
+            "sampleRateHertz": str(sample_rate_hz),
+        }
+        if self._topic:
+            params["topic"] = self._topic
+        if self._folder_id:
+            params["folderId"] = self._folder_id
+
+        with httpx.Client(timeout=self._timeout_seconds) as client:
+            response = client.post(
+                f"{self._api_base}/speech/v1/stt:recognize",
+                headers=self._headers(),
+                params=params,
+                content=pcm_bytes,
+            )
+        response.raise_for_status()
+        payload = response.json()
+        return ASRTranscription(
+            text=str(payload.get("result") or "").strip(),
+            language=language,
+            confidence=None,
+        )
+
+    def _transcribe_v3_async(self, pcm_bytes: bytes, *, sample_rate_hz: int, language: str) -> ASRTranscription:
+        payload = {
+            "content": base64.b64encode(pcm_bytes).decode("ascii"),
+            "recognitionModel": {
+                "model": self._topic or "general",
+                "audioFormat": {
+                    "rawAudio": {
+                        "audioEncoding": "LINEAR16_PCM",
+                        "sampleRateHertz": str(sample_rate_hz),
+                        "audioChannelCount": "1",
+                    }
+                },
+                "textNormalization": {
+                    "textNormalization": "TEXT_NORMALIZATION_ENABLED",
+                    "profanityFilter": False,
+                    "literatureText": False,
+                    "phoneFormattingMode": "PHONE_FORMATTING_MODE_DISABLED",
+                },
+                "languageRestriction": {
+                    "restrictionType": "WHITELIST",
+                    "languageCode": [language],
+                },
+                "audioProcessingType": "FULL_DATA",
+            },
+        }
+        headers = self._headers(content_type="application/json")
+        deadline = time.monotonic() + self._timeout_seconds
+        with httpx.Client(timeout=self._timeout_seconds) as client:
+            response = client.post(
+                f"{self._api_base}/stt/v3/recognizeFileAsync",
+                headers=headers,
+                json=payload,
+            )
+            response.raise_for_status()
+            operation_payload = response.json()
+            operation_id = str(operation_payload.get("id") or "").strip()
+            if not operation_id:
+                raise RuntimeError("Yandex ASR v3 response is missing operation id")
+
+            while not bool(operation_payload.get("done")):
+                if time.monotonic() >= deadline:
+                    raise TimeoutError("Yandex ASR v3 recognition timed out")
+                time.sleep(self._poll_interval_seconds)
+                operation_response = client.get(
+                    f"{self._api_base}/operations/{operation_id}",
+                    headers=headers,
+                )
+                operation_response.raise_for_status()
+                operation_payload = operation_response.json()
+
+            error_payload = operation_payload.get("error")
+            if isinstance(error_payload, dict) and error_payload:
+                raise RuntimeError(str(error_payload.get("message") or error_payload)[:500])
+
+            result_response = client.get(
+                f"{self._api_base}/stt/v3/getRecognition",
+                headers=headers,
+                params={"operationId": operation_id},
+            )
+            result_response.raise_for_status()
+            result_payload = result_response.json()
+
+        return ASRTranscription(
+            text=self._extract_v3_text(result_payload),
+            language=language,
+            confidence=None,
+        )
+
+    @classmethod
+    def _extract_v3_text(cls, payload: object) -> str:
+        if isinstance(payload, list):
+            texts = [cls._extract_v3_text(item) for item in payload]
+            return " ".join(text for text in texts if text).strip()
+        if not isinstance(payload, dict):
+            return ""
+
+        final_refinement = payload.get("finalRefinement")
+        if isinstance(final_refinement, dict):
+            normalized = final_refinement.get("normalizedText")
+            text = cls._extract_alternatives_text(normalized)
+            if text:
+                return text
+
+        for key in ("final", "partial"):
+            text = cls._extract_alternatives_text(payload.get(key))
+            if text:
+                return text
+
+        response = payload.get("response")
+        if isinstance(response, (dict, list)):
+            return cls._extract_v3_text(response)
+        return ""
+
+    @staticmethod
+    def _extract_alternatives_text(payload: object) -> str:
+        if not isinstance(payload, dict):
+            return ""
+        alternatives = payload.get("alternatives")
+        if not isinstance(alternatives, list):
+            return ""
+        texts: list[str] = []
+        for alternative in alternatives:
+            if not isinstance(alternative, dict):
+                continue
+            text = str(alternative.get("text") or "").strip()
+            if text:
+                texts.append(text)
+        return " ".join(texts).strip()
+
+    def transcribe_partial(self, audio_bytes: bytes, *, language_hint: str | None = None) -> ASRTranscription:
+        return self.transcribe(audio_bytes, language_hint=language_hint)
+
+
 class LocalSidecarStreamingASRProvider(StreamingASRProvider):
     name = "local-sidecar"
     supports_streaming = True
@@ -245,10 +551,60 @@ class LocalSidecarStreamingASRProvider(StreamingASRProvider):
             return
 
 
+class YandexSpeechKitBufferedStreamingASRProvider(StreamingASRProvider):
+    name = "yandex-speechkit-buffered"
+    supports_streaming = True
+
+    def __init__(self, *, asr_provider: YandexSpeechKitASRProvider | None = None) -> None:
+        self._asr_provider = asr_provider or YandexSpeechKitASRProvider()
+        self._streams: dict[str, dict[str, object]] = {}
+        self._lock = threading.Lock()
+
+    def open_stream(self, session_id: str, *, language_hint: str | None = None) -> str:
+        stream_id = f"yasr_{uuid.uuid4().hex}"
+        with self._lock:
+            self._streams[stream_id] = {
+                "session_id": str(session_id or "").strip(),
+                "language_hint": str(language_hint or "").strip() or None,
+                "pcm": bytearray(),
+            }
+        return stream_id
+
+    def push_pcm(self, stream_id: str, pcm_8k_chunk: bytes) -> None:
+        if not pcm_8k_chunk:
+            return
+        with self._lock:
+            stream = self._streams.get(stream_id)
+            if stream is None:
+                raise StreamingASRUnavailable("Yandex ASR stream is not active")
+            pcm = stream.get("pcm")
+            if isinstance(pcm, bytearray):
+                pcm.extend(pcm_8k_chunk)
+
+    def poll_partial(self, stream_id: str) -> StreamingASRPartial | None:
+        del stream_id
+        return None
+
+    def finalize(self, stream_id: str) -> ASRTranscription:
+        with self._lock:
+            stream = self._streams.get(stream_id)
+            if stream is None:
+                raise StreamingASRUnavailable("Yandex ASR stream is not active")
+            pcm = bytes(stream.get("pcm") or b"")
+            language_hint = str(stream.get("language_hint") or "").strip() or None
+        return self._asr_provider.transcribe(pcm, language_hint=language_hint)
+
+    def close_stream(self, stream_id: str) -> None:
+        with self._lock:
+            self._streams.pop(stream_id, None)
+
+
 def build_asr_provider(name: str) -> ASRProvider:
     normalized = str(name or "stub").strip().lower()
     if normalized == "openai":
         return OpenAIASRProvider()
+    if normalized in {"yandex", "yandex_speechkit", "speechkit"}:
+        return YandexSpeechKitASRProvider()
     return ASRProvider()
 
 
@@ -256,4 +612,6 @@ def build_streaming_asr_provider(name: str) -> StreamingASRProvider:
     normalized = str(name or "disabled").strip().lower()
     if normalized in {"local_sidecar", "local-sidecar", "sidecar"}:
         return LocalSidecarStreamingASRProvider()
+    if normalized in {"yandex", "yandex_speechkit", "yandex-speechkit", "speechkit"}:
+        return YandexSpeechKitBufferedStreamingASRProvider()
     return StreamingASRProvider()

tests/test_ai_voice_asr_provider.py

@@ -0,0 +1,207 @@
+from __future__ import annotations
+
+from services.ai_voice_runtime_service.audiosocket import pcm16le_to_wav_bytes
+from services.ai_voice_runtime_service.providers import asr as asr_module
+
+
+class _DummyResponse:
+    def __init__(self, json_payload: dict | None = None) -> None:
+        self.content = b"{}"
+        self._json_payload = json_payload or {}
+
+    def raise_for_status(self) -> None:
+        return None
+
+    def json(self) -> dict:
+        return self._json_payload
+
+
+def test_yandex_asr_provider_posts_lpcm_with_tts_credential_fallback(monkeypatch):
+    calls: list[dict] = []
+    pcm = b"\x01\x00" * 160
+    wav_bytes = pcm16le_to_wav_bytes(pcm, sample_rate_hz=8000)
+
+    class _DummyClient:
+        def __init__(self, *, timeout: float) -> None:
+            self.timeout = timeout
+
+        def __enter__(self) -> "_DummyClient":
+            return self
+
+        def __exit__(self, exc_type, exc, tb) -> None:
+            return None
+
+        def post(
+            self,
+            url: str,
+            *,
+            headers: dict[str, str],
+            params: dict[str, str],
+            content: bytes,
+        ) -> _DummyResponse:
+            calls.append(
+                {
+                    "url": url,
+                    "headers": headers,
+                    "params": params,
+                    "content": content,
+                    "timeout": self.timeout,
+                }
+            )
+            return _DummyResponse({"result": "almaty schedule"})
+
+    monkeypatch.delenv("AI_VOICE_ASR_YANDEX_API_KEY", raising=False)
+    monkeypatch.delenv("AI_VOICE_ASR_YANDEX_IAM_TOKEN", raising=False)
+    monkeypatch.setenv("AI_VOICE_TTS_YANDEX_API_KEY", "tts-yandex-key")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_API_BASE", "https://stt.example.test")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_FOLDER_ID", "folder-test")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_TIMEOUT_SECONDS", "5")
+    monkeypatch.setattr(asr_module.httpx, "Client", _DummyClient)
+
+    provider = asr_module.YandexSpeechKitASRProvider()
+    result = provider.transcribe(wav_bytes, language_hint="ru")
+
+    assert result.text == "almaty schedule"
+    assert result.language == "ru-RU"
+    assert len(calls) == 1
+    assert calls[0]["url"] == "https://stt.example.test/speech/v1/stt:recognize"
+    assert calls[0]["headers"]["Authorization"] == "Api-Key tts-yandex-key"
+    assert calls[0]["headers"]["Content-Type"] == "application/octet-stream"
+    assert calls[0]["params"]["lang"] == "ru-RU"
+    assert calls[0]["params"]["format"] == "lpcm"
+    assert calls[0]["params"]["sampleRateHertz"] == "8000"
+    assert calls[0]["params"]["topic"] == "general"
+    assert calls[0]["params"]["folderId"] == "folder-test"
+    assert calls[0]["content"] == pcm
+    assert calls[0]["timeout"] == 5.0
+
+
+def test_yandex_asr_provider_uses_iam_token(monkeypatch):
+    calls: list[dict] = []
+
+    class _DummyClient:
+        def __init__(self, *, timeout: float) -> None:
+            self.timeout = timeout
+
+        def __enter__(self) -> "_DummyClient":
+            return self
+
+        def __exit__(self, exc_type, exc, tb) -> None:
+            return None
+
+        def post(
+            self,
+            url: str,
+            *,
+            headers: dict[str, str],
+            params: dict[str, str],
+            content: bytes,
+        ) -> _DummyResponse:
+            calls.append({"url": url, "headers": headers, "params": params, "content": content})
+            return _DummyResponse({"result": "operator"})
+
+    monkeypatch.delenv("AI_VOICE_ASR_YANDEX_API_KEY", raising=False)
+    monkeypatch.delenv("AI_VOICE_TTS_YANDEX_API_KEY", raising=False)
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_IAM_TOKEN", "iam-token")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_API_BASE", "https://stt.example.test")
+    monkeypatch.setattr(asr_module.httpx, "Client", _DummyClient)
+
+    provider = asr_module.YandexSpeechKitASRProvider()
+    result = provider.transcribe(b"\x02\x00" * 80, language_hint="kk")
+
+    assert result.text == "operator"
+    assert result.language == "kk-KZ"
+    assert calls[0]["headers"]["Authorization"] == "Bearer iam-token"
+    assert calls[0]["params"]["lang"] == "kk-KZ"
+
+
+def test_yandex_asr_provider_supports_kz_v3_async_rest(monkeypatch):
+    calls: list[dict] = []
+
+    class _DummyClient:
+        def __init__(self, *, timeout: float) -> None:
+            self.timeout = timeout
+
+        def __enter__(self) -> "_DummyClient":
+            return self
+
+        def __exit__(self, exc_type, exc, tb) -> None:
+            return None
+
+        def post(self, url: str, *, headers: dict[str, str], json: dict) -> _DummyResponse:
+            calls.append({"method": "POST", "url": url, "headers": headers, "json": json})
+            return _DummyResponse({"id": "operation-1", "done": True})
+
+        def get(
+            self,
+            url: str,
+            *,
+            headers: dict[str, str],
+            params: dict[str, str] | None = None,
+        ) -> _DummyResponse:
+            calls.append({"method": "GET", "url": url, "headers": headers, "params": params or {}})
+            return _DummyResponse(
+                {
+                    "finalRefinement": {
+                        "normalizedText": {
+                            "alternatives": [
+                                {
+                                    "text": "almaty schedule",
+                                }
+                            ]
+                        }
+                    }
+                }
+            )
+
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_API_KEY", "asr-key")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_API_BASE", "https://stt.api.ml.yandexcloud.kz")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_FOLDER_ID", "folder-test")
+    monkeypatch.setenv("AI_VOICE_ASR_YANDEX_POLL_INTERVAL_SECONDS", "0.05")
+    monkeypatch.setattr(asr_module.httpx, "Client", _DummyClient)
+
+    provider = asr_module.YandexSpeechKitASRProvider()
+    result = provider.transcribe(b"\x03\x00" * 80, language_hint="ru")
+
+    assert result.text == "almaty schedule"
+    assert result.language == "ru-RU"
+    assert calls[0]["url"] == "https://stt.api.ml.yandexcloud.kz/stt/v3/recognizeFileAsync"
+    assert calls[0]["headers"]["Authorization"] == "Api-Key asr-key"
+    assert calls[0]["headers"]["Content-Type"] == "application/json"
+    assert calls[0]["headers"]["x-folder-id"] == "folder-test"
+    assert calls[0]["json"]["recognitionModel"]["audioFormat"]["rawAudio"]["sampleRateHertz"] == "8000"
+    assert calls[0]["json"]["recognitionModel"]["languageRestriction"]["languageCode"] == ["ru-RU"]
+    assert calls[1]["url"] == "https://stt.api.ml.yandexcloud.kz/stt/v3/getRecognition"
+    assert calls[1]["params"] == {"operationId": "operation-1"}
+
+
+def test_yandex_buffered_streaming_provider_buffers_pcm_until_finalize():
+    calls: list[dict] = []
+
+    class _FakeYandexASR:
+        def transcribe(self, audio_bytes: bytes, *, language_hint: str | None = None):
+            calls.append({"audio_bytes": audio_bytes, "language_hint": language_hint})
+            return asr_module.ASRTranscription(text="need schedule", language=language_hint, confidence=None)
+
+    provider = asr_module.YandexSpeechKitBufferedStreamingASRProvider(asr_provider=_FakeYandexASR())  # type: ignore[arg-type]
+    stream_id = provider.open_stream("session-1", language_hint="ru")
+
+    provider.push_pcm(stream_id, b"\x10\x00")
+    provider.push_pcm(stream_id, b"\x20\x00")
+
+    assert provider.poll_partial(stream_id) is None
+
+    result = provider.finalize(stream_id)
+    provider.close_stream(stream_id)
+
+    assert result.text == "need schedule"
+    assert calls == [{"audio_bytes": b"\x10\x00\x20\x00", "language_hint": "ru"}]
+
+
+def test_yandex_asr_builders():
+    assert isinstance(asr_module.build_asr_provider("yandex"), asr_module.YandexSpeechKitASRProvider)
+    assert isinstance(asr_module.build_asr_provider("speechkit"), asr_module.YandexSpeechKitASRProvider)
+    assert isinstance(
+        asr_module.build_streaming_asr_provider("yandex_speechkit"),
+        asr_module.YandexSpeechKitBufferedStreamingASRProvider,
+    )