fix(transcription): load PCM WAV via stdlib to skip PyAV decode

[webenefits]

Problem - on Linux (CachyOS)
Transcription after recording can fail with UnicodeDecodeError (e.g. 'ascii' codec can't decode byte …) in av.error.err_check when PyAV decodes FFmpeg error paths as ASCII. This shows up on some locales (e.g. German) when FFmpeg/PyAV surfaces non-ASCII text.

Cause
faster-whisper decodes and resamples via PyAV; FFmpeg error handling can expose non-ASCII strings, and PyAV may decode them incorrectly. In-app recordings are standard PCM WAV.

Solution
For supported .wav files (16-bit PCM, mono or stereo), load audio with the stdlib wave module and NumPy, optionally downmix stereo to mono, linearly resample to feature_extractor.sampling_rate (16 kHz), and pass a float32 np.ndarray into transcribe / BatchedInferencePipeline.transcribe so faster-whisper skips decode_audio. Unsupported WAV variants and all non-WAV inputs keep the previous behavior (file path string → PyAV pipeline unchanged).

Testing

• Synthetic 44.1 kHz mono int16 WAV through the loader (resampled length ~16k at 16 kHz).
• Non-.wav path still yields None → string path unchanged.

Prepared with assistance from Cursor (AI editor). I have tested the changes locally, please still give the patch a careful review before merging.

[gemini-code-assist]

Important

Installation incomplete: to start using Gemini Code Assist, please ask the organization owner(s) to visit the Gemini Code Assist Admin Console and sign the Terms of Services.

[BBC-Esq]

Can you confirm that you're real? I went to the github account and it's barren.

[webenefits]

Hi, yes, I'm real.
Actually, we're not developers, we're a web agency, which is why there isn't much activity on our GitHub account. But we're big fans of open source, and in our opinion, that includes sharing code when you've improved or fixed something for yourself.
Best regards,
Web-Benefits

[BBC-Esq]

Congratulations on your first contribution to this project! Great catch on the PyAV locale issue — clean implementation with a solid fallback. Merged!

[webenefits]

Thank you very much, I'm glad to hear that!