feat: expose condition_on_previous_text parameter to prevent cascading degeneration in long audio

[Benju1]

Problem

Long audio recordings (>30s) with natural speech pauses produce degenerate output:
Repeated commas, truncated transcription, or hallucinated text. This is closely related to #134.

Root Cause

In non-batched mode, faster-whisper concatenates all VAD speech segments into a single audio blob and processes it in 30-second windows (WhisperModel.transcribe(), line audio = np.concatenate(audio_chunks, axis=0)). With condition_on_previous_text=True (the faster-whisper default), degenerate output from one window is fed as context into the next, causing a cascade failure that corrupts all subsequent transcription.

This is a known faster-whisper issue (see SYSTRAN/faster-whisper#843), and it is unlikely to be fixed upstream soon. Speaches can mitigate this today by exposing the condition_on_previous_text parameter that faster-whisper already supports but Speaches does not pass through.

The problem is especially pronounced with:

• Technical speech containing numbers and domain terms (e.g., "14,4 kV")
• Recordings longer than 30 seconds with natural pauses
• German and other non-English languages

Reproduction

Using model TheTobyB/whisper-large-v3-turbo-german-ct2 with German audio:

1. Record ~70 seconds of technical speech with natural pauses between sentences
2. Send to the API:

   curl -X POST "http://localhost:8000/v1/audio/transcriptions" \
     -F "file=@recording.wav" \
     -F "language=de" \
     -F "model=TheTobyB/whisper-large-v3-turbo-german-ct2"

3. Transcription cuts off after ~30s, ending with degenerate output like ",,"

Logs show VAD keeps nearly the entire audio as one segment, and Whisper processes it in 3 x 30s windows:

Processing audio with duration 01:10.490
VAD filter removed 00:02.352 of audio
VAD filter kept the following audio segments: [00:02.352 -> 01:10.490]
Processing segment at 00:00.000
Processing segment at 00:30.000
Processing segment at 01:00.000

Proposed Solution

Expose condition_on_previous_text as a configurable parameter — similar to how #457 made vad_filter configurable:

1. Add condition_on_previous_text to config or as a Form() parameter on the STT endpoints
2. Pass it through to whisper_model.transcribe(..., condition_on_previous_text=...)

Setting it to False makes each 30s window independent, so degeneration in one window stays isolated and subsequent windows recover. This is the minimal, tested fix.

Tradeoff: Slight loss of cross-window context (a sentence spanning a 30s boundary loses prior context). In practice this is negligible compared to losing the entire second half of a transcription.

Workarounds

Until this is configurable, two workarounds exist:

Approach	How	Tradeoff
condition_on_previous_text=False	Volume-mount patched stt.py	Minimal accuracy loss at window boundaries
WHISPER__USE_BATCHED_MODE=true	Config change, no code patch	More word errors on technical vocabulary (each VAD segment loses all prior context)

Both produce complete transcriptions. condition_on_previous_text=False delivers better word accuracy on technical content.

Environment

• speaches latest-cpu (digest sha256:21e3df06d842...)
• Model: TheTobyB/whisper-large-v3-turbo-german-ct2
• Non-batched mode (default)