Voice-to-Text Transcription API

System design

 Client ──POST /transcribe (audio)──►
   1. Validate      extension · MIME type · size
   2. Store         temporary file
   3. Normalise     FFmpeg → mono 16 kHz WAV
   4. Transcribe    faster-whisper
   5. Segment       [{ start, end, text }]
   6. Clean up      delete temp files
 ◄── JSON { text, segments, duration }

Key decisions

• Normalise before inference — one audio format in means predictable accuracy and speed.
• Validate at the edge — reject bad files before they reach FFmpeg or the model.
• faster-whisper (CTranslate2) for CPU-friendly inference.
• Clear next steps documented: a job queue for long files, streaming results, and GPU workers for scale.