TTS streaming fails

[fawzib]

i am getting the below error when i use this HA plugin: https://github.com/sfortis/openai_tts/

url i used: http://192.168.1.60:8000/v1/realtime
i tried other urls as well and all failed.

2025-12-08 00:06:25,658:ERROR:speaches.routers.stt:transcription_response_to_http_response:105:Unexpected streaming transcription response type: <class 'openai.types.audio.transcription.Transcription'>
INFO:     192.168.1.236:57820 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK
INFO:     192.168.1.236:57830 - "POST /v1/realtime HTTP/1.1" 422 Unprocessable Entity
2025-12-08 00:06:43,286:DEBUG:speaches.dependencies:audio_file_dependency:78:Decoding audio file: whisper_audio.wav, content_type: audio/wav, header: Headers({'content-type': 'audio/wav', 'content-disposition': 'form-data; name="file"; filename="whisper_audio.wav"'}), size: 421164
2025-12-08 00:06:43,385:DEBUG:speaches.dependencies:audio_file_dependency:92:Decoded 13.16s of audio in 0.09933s (RTF: 0.007547672342057524)
2025-12-08 00:06:43,385:DEBUG:asyncio:__init__:64:Using selector: EpollSelector
2025-12-08 00:06:43,394:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:silero_vad_v5 model already loaded
2025-12-08 00:06:43,394:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for silero_vad_v5, self.ref_count=1
2025-12-08 00:06:43,453:DEBUG:speaches.executors.silero_vad_v5:__call__:135:VAD model inference took 0.0588s
2025-12-08 00:06:43,454:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for silero_vad_v5, self.ref_count=0
2025-12-08 00:06:43,454:INFO:speaches.executors.shared.base_model_manager:_decrement_ref:98:Model silero_vad_v5 is idle, not unloading
2025-12-08 00:06:43,454:DEBUG:speaches.executors.silero_vad_v5:get_speech_timestamps:303:VAD processing took 0.0602s for 13.16s audio
2025-12-08 00:06:43,454:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:Systran/faster-distil-whisper-medium.en model already loaded
2025-12-08 00:06:43,454:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:81:Model was set to expire in 21600s, cancelling
2025-12-08 00:06:43,454:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for Systran/faster-distil-whisper-medium.en, self.ref_count=1
2025-12-08 00:06:43,454:INFO:faster_whisper:transcribe:391:Processing audio with duration 00:13.160
2025-12-08 00:06:43,455:INFO:faster_whisper:transcribe:428:VAD filter removed 00:00.000 of audio
2025-12-08 00:06:43,823:INFO:speaches.executors.whisper:handle_non_streaming_transcription_request:177:Transcribed 13.16 seconds of audio in 0.3682780009985436 seconds
2025-12-08 00:06:43,823:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for Systran/faster-distil-whisper-medium.en, self.ref_count=0
2025-12-08 00:06:43,823:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:91:Model Systran/faster-distil-whisper-medium.en is idle, scheduling offload in 21600s
2025-12-08 00:06:43,823:ERROR:speaches.routers.stt:transcription_response_to_http_response:105:Unexpected streaming transcription response type: <class 'openai.types.audio.transcription.Transcription'>
INFO:     192.168.1.236:36448 - "POST /v1/audio/transcriptions HTTP/1.1" 200 OK

[maglat]

I have no solution for this, but all STT + TTS services I include to Home Assistant via https://github.com/roryeckel/wyoming_openai

For STT I use Speaches, but for TTS I use Chatterbox (https://github.com/devnen/Chatterbox-TTS-Server)

This Chatterbox server + wyoming_openai works very reliable with streaming.
The solution you mentioned above, cancel long texts very often.

[Starystars67]

I have been trying to get STT working properly and though I can get it to work I am seeing issues with it trying to generate a response back even though I set session.turn_detection.create_response = false.

Would you mind sharing how you handle it?

https://github.com/roryeckel/wyoming_openai

[ordex]

I am having the same output as @fawzib

2026-01-27 13:28:42,858:DEBUG:speaches.dependencies:audio_file_dependency:89:Decoding audio file: 4969dade-8fc6-418f-b041-899937150ace.mp3, content_type: None, header: Headers({'content-disposition': 'form-data; name="file"; filename="4969dade-8fc6-418f-b041-899937150ace.mp3"'}), size: 13677
2026-01-27 13:28:42,924:DEBUG:speaches.dependencies:audio_file_dependency:103:Decoded 1.6335s of audio in 0.06585s (RTF: 0.04031378941175439)
2026-01-27 13:28:42,925:DEBUG:asyncio:__init__:64:Using selector: EpollSelector
2026-01-27 13:28:42,926:DEBUG:speaches.hf_utils:passes_filter:41:Model ID 'istupakov/parakeet-tdt-0.6b-v3-onnx' does not match filter library 'ctranslate2': {'base_model': ['nvidia/parakeet-tdt-0.6b-v3'], 'language': ['en', 'es', 'fr', 'de', 'bg', 'hr', 'cs', 'da', 'nl', 'et', 'fi', 'el', 'hu', 'it', 'lv', 'lt', 'mt', 'pl', 'pt', 'ro', 'sk', 'sl', 'sv', 'ru', 'uk'], 'license': 'cc-by-4.0', 'pipeline_tag': 'automatic-speech-recognition'}
2026-01-27 13:28:42,926:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:silero_vad_v5 model already loaded
2026-01-27 13:28:42,926:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for silero_vad_v5, self.ref_count=1
2026-01-27 13:28:42,929:DEBUG:speaches.executors.silero_vad_v5:__call__:138:VAD model inference took 0.0029s
2026-01-27 13:28:42,929:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for silero_vad_v5, self.ref_count=0
2026-01-27 13:28:42,929:INFO:speaches.executors.shared.base_model_manager:_decrement_ref:98:Model silero_vad_v5 is idle, not unloading
2026-01-27 13:28:42,929:DEBUG:speaches.executors.silero_vad_v5:get_speech_timestamps:307:VAD processing took 0.0031s for 1.63s audio
2026-01-27 13:28:42,929:DEBUG:speaches.executors.shared.base_model_manager:load_model:138:istupakov/parakeet-tdt-0.6b-v3-onnx model already loaded
2026-01-27 13:28:42,929:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:81:Model was set to expire in 300s, cancelling
2026-01-27 13:28:42,929:DEBUG:speaches.executors.shared.base_model_manager:_increment_ref:83:Incremented ref count for istupakov/parakeet-tdt-0.6b-v3-onnx, self.ref_count=1
2026-01-27 13:28:43,039:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:88:Decremented ref count for istupakov/parakeet-tdt-0.6b-v3-onnx, self.ref_count=0
2026-01-27 13:28:43,039:DEBUG:speaches.executors.shared.base_model_manager:_decrement_ref:91:Model istupakov/parakeet-tdt-0.6b-v3-onnx is idle, scheduling offload in 300s
2026-01-27 13:28:43,039:ERROR:speaches.routers.stt:transcription_response_to_http_response:105:Unexpected streaming transcription response type: <class 'openai.types.audio.transcription.Transcription'>

when issuing a request via Open-WebUI, however, it seems to be working as I am getting the proper text entered in the Open-WebUI prompt box.

Any hint as to why this is happening and what it may be affecting?

[gitjfmd]

Looking at master, the logger.error(...) at the top of transcription_response_to_http_response in src/speaches/routers/stt.py (around line 105) fires unconditionally, before the isinstance dispatch runs:

def transcription_response_to_http_response(res):
    logger.error(f"Unexpected streaming transcription response type: {type(res)}")
    if isinstance(res, tuple): ...
    elif isinstance(res, (Transcription, TranscriptionVerbose)): ...
    else: ...

The dispatch below correctly handles Transcription / TranscriptionVerbose, which is why the HTTP response still comes back 200. The log line just gets emitted on every call regardless of type. Same symptom @ordex described above with parakeet, we hit it with deepdml/faster-whisper-large-v3-turbo-ct2 on non-streaming POST /v1/audio/transcriptions.

Fix is to move the log + raise into an actual fallthrough branch instead of running it at the top:

if isinstance(res, tuple): ...
if isinstance(res, (Transcription, TranscriptionVerbose)): ...
if isinstance(res, Generator): ...
logger.error(f"Unexpected transcription response type: {type(res)}")
raise TypeError(f"Unexpected transcription response type: {type(res)}")

Happy to PR.