Audio Normalize — Voice Balancing for Multi-Speaker Recordings

Balances voices in a conversation recording where some speakers are far from the microphone. All processing runs on-device — no audio ever leaves the Mac. Works with French and English recordings.

When to Use

• Some voices in the recording are too quiet (far mic distance)
• Preparing an .m4a before Apple Intelligence transcription (re-inject on iPhone/Mac) or before running the meeting-transcribe skill
• Want a local on-device transcript without any cloud service
• Pre-processing a Voice Memo before sharing with teammates

What It Does

Three-stage ffmpeg chain:

Output: <original>_normalized.m4a — safe to use with any transcription tool.

Procedure

Step 0 — Confirm Input

Ask the user for the path to the audio file if not provided. Confirm output location (default: same folder, _normalized suffix).

Step 1 — Check Environment

# ffmpeg — required for all processing
which ffmpeg || echo "Install: brew install ffmpeg"

# whisper-cli — only needed for local transcription
which whisper-cli || echo "Install: brew install whisper-cpp"

If ffmpeg is missing, stop and ask the user to install it first.

Step 2 — Run Normalization

chmod +x ./scripts/normalize.sh

# Basic — normalize only (then re-inject on iPhone or use Apple Intelligence)
./scripts/normalize.sh path/to/recording.m4a

# With local on-device transcription (whisper.cpp, no internet)
./scripts/normalize.sh path/to/recording.m4a --transcribe

# Specify language explicitly (faster, slightly more accurate)
./scripts/normalize.sh path/to/recording.m4a --transcribe --lang fr
./scripts/normalize.sh path/to/recording.m4a --transcribe --lang en

The script prints loudness before/after so you can see the effect.

Step 3 — Transcription Options

Present the user with the two paths:

Option A — Apple Intelligence (recommended for iPhone integration)

1. AirDrop _normalized.m4a to iPhone
2. Open in Voice Memos → tap the recording → share to Notes or use transcript button
3. Or on Mac: open in QuickTime Player → Edit menu → Transcribe

Option B — Local whisper.cpp (best pure-local quality, no Apple account)

Use --transcribe flag above, or set up once:

brew install whisper-cpp
# Download best model for M1 Pro (quantized, ~1 GB, fast on Metal GPU)
mkdir -p ~/.cache/whisper
curl -L "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-q5_0.bin" \
     -o ~/.cache/whisper/ggml-large-v3-q5_0.bin
# Model lands in ~/.cache/whisper/ — script finds it automatically

whisper.cpp uses Apple Metal GPU on M1 — transcribes ~10 min audio in ~30 s. Supports --lang auto for mixed French/English sessions.

Option C — meeting-transcribe skill (adds speaker diarization)

If you need to know who said what, feed the normalized output into the meeting-transcribe skill:

/meeting-transcribe path/to/recording_normalized.m4a

Step 4 — Verify Output

• Output file should be ~same size as input (AAC 128k)
• Listen to a quiet passage: far voices should be clearly audible
• Loudness printed at end should be near −16 LUFS

Tuning (Advanced)

Default compressor settings in normalize.sh work well for most recordings. If results are over-compressed (voices sound pumping/unnatural), reduce ratio:

# In normalize.sh, find:
acompressor=threshold=-32dB:ratio=5:attack=5:release=80:makeup=5dB
# Reduce ratio to 3 or 4 for a gentler touch:
acompressor=threshold=-32dB:ratio=3:attack=5:release=80:makeup=3dB

Files