digiops/realtime_voice_service
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'refs/remotes/origin/main'

Browse Source
This commit is contained in:
Magzhan Zhumabayev
2026-05-01 18:08:56 +05:00
parent 5dda1e10a3 d4d8555bd9
commit 676d7155bd
10 changed files with 87 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.env.example
+1 -1
View File
@@ -2,7 +2,7 @@ REALTIME_VOICE_HTTP_HOST=0.0.0.0
REALTIME_VOICE_HTTP_PORT=8000
REALTIME_VOICE_AUDIOSOCKET_HOST=0.0.0.0
REALTIME_VOICE_AUDIOSOCKET_PORT=9092
REALTIME_VOICE_SAMPLE_RATE_HZ=8000
REALTIME_VOICE_SAMPLE_RATE_HZ=16000
LLM_PROVIDER=openai
core/__init__.py
+19 -3
View File
@@ -1,4 +1,20 @@
from realtime_voice_service.core.session import CallSession, SessionState
from realtime_voice_service.core.vad import BaseVAD, SileroVADDetector, VADFrameResult
__all__ = ["BaseVAD", "CallSession", "SessionState", "SileroVADDetector", "VADFrameResult"]
def __getattr__(name: str):
if name in {"CallSession", "SessionState"}:
from realtime_voice_service.core.session import CallSession, SessionState
return {
"CallSession": CallSession,
"SessionState": SessionState,
}[name]
if name in {"BaseVAD", "SileroVADDetector", "VADFrameResult"}:
from realtime_voice_service.core.vad import BaseVAD, SileroVADDetector, VADFrameResult
return {
"BaseVAD": BaseVAD,
"SileroVADDetector": SileroVADDetector,
"VADFrameResult": VADFrameResult,
}[name]
raise AttributeError(name)
core/session.py
+1 -7
View File
@@ -2153,13 +2153,7 @@ class CallSession:
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S_%f")
safe_session_id = re.sub(r"[^A-Za-z0-9_.-]+", "_", self.session_id)
filename = os.path.join(dump_dir, f"utterance_{safe_session_id}_{timestamp}.wav")
dump_sample_rate_hz = 8000
if self.transport.sample_rate_hz != dump_sample_rate_hz:
LOGGER.warning(
"realtime session %s audio dump writing raw bytes with 8000Hz header while transport sample_rate=%s",
self.session_id,
self.transport.sample_rate_hz,
)
dump_sample_rate_hz = self.transport.sample_rate_hz
with wave.open(filename, "wb") as wav_file:
wav_file.setnchannels(1)
core/vad.py
+1 -1
View File
@@ -54,7 +54,7 @@ class SileroVADDetector(BaseVAD):
def __init__(
self,
*,
sample_rate_hz: int = 8000,
sample_rate_hz: int = 16000,
threshold: float = 0.5,
negative_threshold: float | None = None,
speech_end_silence_ms: int = 1600,
main.py
+8 -1
View File
@@ -72,7 +72,9 @@ def _audiosocket_port() -> int:
def _sample_rate_hz() -> int:
configured = max(_int_env("REALTIME_VOICE_SAMPLE_RATE_HZ", 16000), 1)
return configured if configured in {8000, 16000, 24000} else 16000
if configured != 16000:
LOGGER.warning("REALTIME_VOICE_SAMPLE_RATE_HZ=%s is not allowed; using 16000 Hz", configured)
return 16000
def _http_host() -> str:
@@ -148,6 +150,10 @@ class RealtimeVoiceService:
def active_session_count(self) -> int:
return len(self._active_sessions)
@property
def sample_rate_hz(self) -> int:
return self._sample_rate_hz
async def start(self) -> None:
LOGGER.info("realtime voice service starting")
await self._audiosocket_server.start()
@@ -285,6 +291,7 @@ async def health() -> dict[str, object]:
return {
"status": "ok",
"service": "realtime-voice-service",
"sample_rate_hz": service.sample_rate_hz,
"active_sessions": service.active_session_count,
"audiosocket_port": _audiosocket_port(),
}
providers/factory.py
+1 -1
View File
@@ -68,7 +68,7 @@ def _build_single_stt_provider(
) -> BaseSTT:
normalized = provider.lower().strip()
if normalized in {"openai", "whisper", "openai_whisper"}:
# Whisper accepts an 8 kHz WAV container; keep the exact AudioSocket PCM, no resampling.
# Whisper accepts the configured WAV container; keep the exact AudioSocket PCM, no resampling.
return OpenAISTT(input_sample_rate_hz=input_sample_rate_hz)
if normalized in {"elevenlabs", "eleven_labs", "scribe"}:
return ElevenLabsSTT(
providers/stt.py
+2 -2
View File
@@ -1010,8 +1010,8 @@ class YandexSpeechKitSTT(BaseSTT):
api_base: str | None = None,
language: str | None = None,
topic: str | None = None,
input_sample_rate_hz: int = 8000,
target_sample_rate_hz: int = 8000,
input_sample_rate_hz: int = 16000,
target_sample_rate_hz: int = 16000,
timeout_seconds: float | None = None,
) -> None:
self._api_key = str(api_key if api_key is not None else _yandex_api_key()).strip()
providers/stt_openai.py
+1 -1
View File
@@ -79,7 +79,7 @@ class OpenAISTT(BaseSTT):
api_key: str | None = None,
model: str | None = None,
base_url: str | None = None,
input_sample_rate_hz: int = 8000,
input_sample_rate_hz: int = 16000,
prompt: str | None = None,
language: str | None = None,
timeout_seconds: float | None = None,
providers/tts.py
+6 -3
View File
@@ -119,11 +119,15 @@ class ElevenLabsTTS(BaseTTS):
str(output_format or os.getenv("ELEVENLABS_TTS_OUTPUT_FORMAT", "pcm_16000")).strip().lower()
or "pcm_16000"
)
requested_sample_rate_hz = _sample_rate_from_pcm_format(requested_output_format)
self._target_sample_rate_hz = int(
target_sample_rate_hz
if target_sample_rate_hz is not None
else (_sample_rate_from_pcm_format(requested_output_format) or 16000)
else (requested_sample_rate_hz or 16000)
)
if self._target_sample_rate_hz == 8000:
LOGGER.warning("ElevenLabs TTS target sample rate 8000Hz is disabled; using 16000Hz")
self._target_sample_rate_hz = 16000
self._output_format = self._resolve_provider_output_format(
requested_output_format=requested_output_format,
target_sample_rate_hz=self._target_sample_rate_hz,
@@ -381,8 +385,7 @@ class ElevenLabsTTS(BaseTTS):
normalized = requested_output_format.strip().lower() or "pcm_16000"
if normalized == "pcm_8000":
LOGGER.warning(
"ElevenLabs TTS does not provide reliable PCM16 8kHz streaming; requesting pcm_16000 "
"and resampling to %sHz locally",
"ElevenLabs TTS PCM16 8kHz output is disabled; requesting pcm_16000 for %sHz target",
target_sample_rate_hz,
)
return "pcm_16000"
transports/audiosocket.py
+47 -6
View File
@@ -17,7 +17,22 @@ LOGGER = logging.getLogger("uvicorn.error")
AUDIO_SOCKET_PACKET_HANGUP = 0x00
AUDIO_SOCKET_PACKET_UUID = 0x01
AUDIO_SOCKET_PACKET_DTMF = 0x03
AUDIO_SOCKET_PACKET_PCM16 = 0x10
_AUDIO_SOCKET_PCM16_PACKET_TYPES_BY_SAMPLE_RATE_HZ: dict[int, int] = {
8000: 0x10,
12000: 0x11,
16000: 0x12,
24000: 0x13,
32000: 0x14,
44100: 0x15,
48000: 0x16,
96000: 0x17,
192000: 0x18,
}
_AUDIO_SOCKET_SAMPLE_RATE_HZ_BY_PCM16_PACKET_TYPE: dict[int, int] = {
packet_type: sample_rate_hz
for sample_rate_hz, packet_type in _AUDIO_SOCKET_PCM16_PACKET_TYPES_BY_SAMPLE_RATE_HZ.items()
}
SUPPORTED_AUDIO_SOCKET_SAMPLE_RATES_HZ = frozenset(_AUDIO_SOCKET_PCM16_PACKET_TYPES_BY_SAMPLE_RATE_HZ)
def _float_env(name: str, default: float) -> float:
@@ -43,8 +58,20 @@ def encode_packet(packet_type: int, payload: bytes = b"") -> bytes:
return struct.pack("!BH", packet_type & 0xFF, len(payload)) + payload
def encode_audio_packet(pcm_bytes: bytes) -> bytes:
return encode_packet(AUDIO_SOCKET_PACKET_PCM16, pcm_bytes)
def pcm16_packet_type_for_sample_rate(sample_rate_hz: int) -> int:
try:
return _AUDIO_SOCKET_PCM16_PACKET_TYPES_BY_SAMPLE_RATE_HZ[int(sample_rate_hz)]
except KeyError as exc:
supported = ", ".join(str(rate) for rate in sorted(SUPPORTED_AUDIO_SOCKET_SAMPLE_RATES_HZ))
raise ValueError(f"unsupported AudioSocket PCM16 sample rate {sample_rate_hz}; supported: {supported}") from exc
def sample_rate_for_pcm16_packet_type(packet_type: int) -> int | None:
return _AUDIO_SOCKET_SAMPLE_RATE_HZ_BY_PCM16_PACKET_TYPE.get(packet_type)
def encode_audio_packet(pcm_bytes: bytes, *, sample_rate_hz: int = 16000) -> bytes:
return encode_packet(pcm16_packet_type_for_sample_rate(sample_rate_hz), pcm_bytes)
async def read_packet(
@@ -78,6 +105,7 @@ class AudioSocketTransport(BaseMediaTransport):
)
self._reader = reader
self._writer = writer
self._pcm_packet_type = pcm16_packet_type_for_sample_rate(self.sample_rate_hz)
self._read_timeout_seconds = max(read_timeout_seconds, 1.0)
self._closed = False
self._rx_audio_packet_count = 0
@@ -166,7 +194,20 @@ class AudioSocketTransport(BaseMediaTransport):
self._rx_audio_bytes,
)
return None
if packet_type == AUDIO_SOCKET_PACKET_PCM16:
packet_sample_rate_hz = sample_rate_for_pcm16_packet_type(packet_type)
if packet_sample_rate_hz is not None:
if packet_sample_rate_hz != self.sample_rate_hz:
LOGGER.error(
"AudioSocket sample rate mismatch: session=%s expected_sample_rate=%s expected_packet_type=0x%02x "
"received_sample_rate=%s received_packet_type=0x%02x payload_bytes=%s",
self.transport_id,
self.sample_rate_hz,
self._pcm_packet_type,
packet_sample_rate_hz,
packet_type,
len(payload),
)
return None
self._rx_audio_packet_count += 1
self._rx_audio_bytes += len(payload)
self._track_rx_audio_level(payload)
@@ -186,7 +227,7 @@ class AudioSocketTransport(BaseMediaTransport):
continue
self._rx_ignored_packet_count += 1
LOGGER.warning(
"AudioSocket unknown packet ignored: session=%s packet_type=%s payload_bytes=%s",
"AudioSocket unknown packet ignored: session=%s packet_type=0x%02x payload_bytes=%s",
self.transport_id,
packet_type,
len(payload),
@@ -196,7 +237,7 @@ class AudioSocketTransport(BaseMediaTransport):
async def _send_frame(self, frame: bytes) -> None:
if self._closed or self._writer.is_closing():
return
self._writer.write(encode_audio_packet(frame))
self._writer.write(encode_audio_packet(frame, sample_rate_hz=self.sample_rate_hz))
await self._writer.drain()
async def close(self) -> None: