Transcribing Voice Memos with whisper.cpp

Transcribing Voice Memos with whisper.cpp on macOS

Ever recorded a long meeting on your iPhone with the Voice Memos app and wished you could get a clean text transcript without sending the audio to some cloud service? Here’s how to do it entirely locally on your Mac using whisper.cpp - free, private, and surprisingly fast on Apple Silicon.

---

What You Need

• A Mac with Apple Silicon (M1 or newer)
• Xcode Command Line Tools
• Homebrew
• The recording exported from Voice Memos (it comes out as a .qta file when dragged off)

---

Step 1 - Build whisper.cpp

Clone the repo and compile it. The make command handles everything and takes under a minute on Apple Silicon.¹

git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp
make -j

If make isn’t found, install Xcode Command Line Tools first: xcode-select --install

The compiled binary ends up at ./build/bin/whisper-cli in newer versions of the project.¹

---

Step 2 - Download a Model

All Whisper models are free to download - no paywall, unlike GUI apps. For meeting transcription, large-v3-turbo is a great balance of speed and accuracy. If you need maximum accuracy and don’t mind waiting longer, use large-v3.²

# Fast and accurate
bash models/download-ggml-model.sh large-v3-turbo

# Maximum accuracy
bash models/download-ggml-model.sh large-v3

---

Step 3 - Convert Your Audio File

Voice Memos exports a .qta file. FFmpeg handles it natively - just convert it to a 16kHz mono WAV, which is the format Whisper expects.³

ffmpeg -i boardmeeting.qta -acodec pcm_s16le -ac 1 -ar 16000 meeting.wav

If you prefer, you can also export directly as M4A from the Voice Memos app on Mac (right-click → Share → Save to Files), which works identically.

---

Step 4 - Transcribe

./build/bin/whisper-cli -m ./models/ggml-large-v3-turbo.bin \
  -l en \
  -f meeting.wav \
  --output-txt --output-vtt

Use -l fr for French, or any other ISO 639-1 language code. The --output-txt flag produces a clean plain-text file, while --output-vtt adds timestamps - you can use both at the same time.

For feeding the transcript into an LLM afterward, use the .txt file - the VTT timestamp noise wastes tokens without adding value for summarization.

---

Fixing Hallucination Loops

On long recordings or sections with silence/background noise, Whisper can get stuck repeating the same phrase endlessly:

[00:06:00.020 --> 00:06:01.020]   So we can wait for the next question.
[00:06:01.020 --> 00:06:02.020]   So we can wait for the next question.
[00:06:02.020 --> 00:06:03.020]   So we can wait for the next question.
...

This is a known Whisper behavior triggered by silence or low-energy audio. There are two complementary fixes.¹

Fix 1 - Strip Silence with FFmpeg

Pre-process the audio to remove silent sections before Whisper ever sees them:

ffmpeg -i meeting.wav \
  -af "silenceremove=stop_periods=-1:stop_duration=2:stop_threshold=-35dB" \
  meeting_clean.wav

Adjust -35dB to -40dB for a more conservative threshold (removes only very quiet sections) or -30dB to be more aggressive.

Fix 2 - Raise Detection Thresholds

Add entropy and probability thresholds to tell Whisper to discard repetitive or low-confidence output:

./build/bin/whisper-cli -m ./models/ggml-large-v3-turbo.bin \
  -l en \
  -f meeting_clean.wav \
  --output-txt --output-vtt \
  --entropy-thold 3.4 \
  --no-speech-thold 0.8 \
  --logprob-thold -0.5

Using both fixes together (cleaned audio + thresholds) is the most robust approach for board meetings and other long recordings.

---

Handling Very Long Recordings (4+ Hours)

Whisper can degrade on very long files. The most reliable solution is to split the audio into chunks and transcribe each one separately:¹

# Split into 30-minute chunks
ffmpeg -i meeting_clean.wav \
  -f segment -segment_time 1800 -c copy chunk_%03d.wav

# Transcribe all chunks
for f in chunk_*.wav; do
  ./build/bin/whisper-cli -m ./models/ggml-large-v3-turbo.bin \
    -l en -f "$f" \
    --output-txt \
    --entropy-thold 3.4 \
    --no-speech-thold 0.8 \
    --logprob-thold -0.5
done

# Merge transcripts
cat chunk_*.wav.txt > full_transcript.txt

This also has the benefit that a bad section in one chunk won’t affect the rest of the transcription.

---

Model Comparison

Model	Speed (Apple Silicon)	Accuracy	Best For
base	Very fast	Low	Quick drafts, testing
small	Fast	Medium	Short, clear recordings
large-v3-turbo	~4x real-time	High	Most meetings
large-v3	~1x real-time	Highest	Maximum accuracy needed

For a 1-hour meeting, large-v3-turbo takes roughly 15 minutes while large-v3 can take closer to 60 minutes.

---

What’s Next

Once you have the .txt transcript, you can pipe it straight into a local LLM (Ollama works great) to generate formal meeting minutes:

ollama run llama3 "Generate formal meeting minutes from this transcript:

$(cat full_transcript.txt)"

Everything in this workflow runs 100% locally - no API keys, no cloud uploads, no subscriptions.

---

References

1. ggml-org/whisper.cpp: Port of OpenAI’s Whisper model in C/C++ ↩ ↩² ↩³ ↩⁴

2. models/download-ggml-model.cmd - whisper.cpp - Hugging Face ↩

3. Running Whisper on an M1 Mac to transcribe audio data locally ↩