[#13] Add audio preprocessing for uploaded recordings

[julien731]

Story: #13

Summary

Add an audio preprocessing pipeline that automatically cleans uploaded audio before transcription. The pipeline applies three stages: 80Hz high-pass filter (removes rumble), conservative noise reduction via noisereduce (prop_decrease=0.75), and loudness normalization to -23 LUFS via pyloudnorm. Preprocessing is enabled by default and can be toggled off per upload via a new checkbox in the upload form.

Key changes:

• New backend/services/audio_preprocessor.py service with the three-stage pipeline
• preprocess_audio boolean field added to MeetingMetadata (defaults to True)
• Upload endpoint accepts preprocess_audio form parameter
• Transcription pipeline calls preprocessor before WhisperX, using a working copy (original preserved)
• Preprocessed file is cleaned up in a finally block after transcription
• Upload form includes an "Audio preprocessing" checkbox with description
• New dependencies: noisereduce>=3.0.0, pyloudnorm>=0.1.1

Approach

Preprocessing is implemented as a separate service module to keep the transcriber focused on transcription. The preprocessed audio is saved as a WAV working copy alongside the original file (preserving the original per T3). Conservative noise reduction settings are used because Whisper was trained on noisy data (BR-3/T1). The preprocessed copy is fed to both transcription and diarization (T2).

Verification

• 119 tests passing (pytest -x -q)
• Unit tests cover: preprocessor output path, original preservation, valid WAV output, stereo-to-mono conversion, spec constants
• Integration tests cover: preprocess_audio defaults to true, can be disabled via form field
• Ruff check and format pass
• Architect review: no critical or major issues

[julien731]

QA Confidence Verdict

What Was Verified

Truths:

Truth	Status	Method
T1: Noise reduction uses prop_decrease=0.75	PASS	Code inspection: NOISE_PROP_DECREASE = 0.75 constant used in nr.reduce_noise(). Unit test test_constants_match_spec asserts this.
T2: Preprocessing applied to both transcription and diarization	PASS	Code inspection: audio_path is reassigned to preprocessed path before any WhisperX calls. All downstream steps (transcribe, align, diarize) use the same variable.
T3: Original uploaded file is preserved; preprocessing creates a working copy	PASS	Code inspection: preprocessed file saved as audio_preprocessed.wav alongside original. Unit test test_preserves_original_file verifies original is untouched. Preprocessed copy is cleaned up in finally block.

Acceptance Criteria:

AC	Status	Method
High-pass filter, noise reduction, loudness normalization applied	PASS	Code inspection: all three stages implemented in audio_preprocessor.py (80Hz Butterworth, noisereduce, pyloudnorm to -23 LUFS)
Preprocessing enabled by default	PASS	preprocess_audio: bool = True in schema. Form default is "true". Integration test confirms.
Upload form includes toggle to disable preprocessing	PASS	Code inspection: checkbox with id preprocess-checkbox, checked by default, wired through api.js to backend.
Conservative noise reduction does not degrade clean audio	PASS	Enforced by prop_decrease=0.75 and stationary=False settings.
preprocess_audio persisted in MeetingMetadata	PASS	Pydantic field added. Integration tests verify persistence for both true and false values.
Dependencies: noisereduce>=3.0.0, pyloudnorm>=0.1.1	PASS	Both added to requirements.txt with correct version constraints.

Tests: All 30 tests pass (5 unit tests for preprocessor, 2 integration tests for preprocess_audio persistence).

What Needs Human Eyes

• Visual appearance of the preprocessing checkbox in the upload form (alignment, spacing, dark/light theme)
• Preprocessing hint text readability

Risk Areas

• No test exercises the full transcription pipeline with preprocessing enabled (transcriber tests set preprocess_audio=False). The integration between preprocessor and transcriber is verified by code inspection only.
• No test verifies that the preprocessed file is cleaned up after transcription completes.

Suggested QA Focus

• Quick glance: Upload form checkbox appearance in both themes
• Functional test: Upload a file with preprocessing enabled and disabled, verify both complete successfully
• Estimated effort: 10 minutes manual testing