As a user, I can have noisy audio automatically cleaned before transcription so that accuracy improves

[julien731]

User Story

As a user uploading a meeting recorded in a large room, I can have the system automatically clean up the audio before transcription so that distant speakers and background noise don't degrade accuracy.

Spec: docs/specs/transcription-quality-improvements.md — US-1

Size: M

Acceptance Criteria

• Uploaded audio undergoes preprocessing before transcription: high-pass filter, noise reduction, and loudness normalization
• Preprocessing is enabled by default
• The upload form includes a toggle/checkbox to disable preprocessing per upload
• Given a clean, close-mic recording with preprocessing enabled, then audio quality is not degraded (conservative noise reduction)
• The preprocess_audio field is persisted in MeetingMetadata
• New dependencies: noisereduce (>=3.0.0) and pyloudnorm (>=0.1.1)

Truths

• T1: Noise reduction uses conservative settings (prop_decrease=0.75) — Whisper was trained on noisy data and aggressive removal hurts accuracy (BR-3)
• T2: Preprocessing is applied to the audio fed to both transcription and diarization (simplest approach, pending OQ#1)
• T3: The original uploaded audio file is preserved; preprocessing creates a working copy

Business Rules

• BR-3: Audio preprocessing uses conservative noise reduction (prop_decrease=0.75). Aggressive noise removal can strip speech harmonics and hurt accuracy.

Edge Cases

• Audio is already clean (close-mic) → conservative preprocessing should not degrade quality; user can disable per upload
• Preprocessing disabled → audio goes directly to transcription pipeline as today

Dependencies

None