Transcribing MP3s with whisper-cpp on macOS#

I asked on Twitter for tips about running Whisper transcriptions in the CLI on my Mac. Werner Robitza pointed me to Homebrew’s whisper-cpp formula, and when I complained that it didn’t have quite enough documentation for me to know how to use it Werner got a PR accepted adding the missing details.

Here’s my recipe for using it to transcribe an MP3 file.

Install whisper-cpp:
Terminal window
```
1
brew install whisper-cpp
```
This gave me a /opt/homebrew/bin/whisper-cpp, added to my PATH as whisper-cpp
Download a Whisper model file. These are available on Hugging Face - there are a bunch of options, I decided to go for ggml-large-v3-q5_0.bin (direct download link, 1GB) because it looked like it might offer right balance of file size to quality.
Terminal window
```
1
curl -o ggml-large-v3-q5_0.bin -L \
2
  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-q5_0.bin?download=true'
```
Convert the MP3 file to the 16khz WAV file needed by Whisper:
Terminal window
```
1
ffmpeg -i input.mp3 -ar 16000 input.wav
```

Run the transcription:

1
whisper-cpp -m ggml-large-v3-q5_0.bin input.wav --output-txt

This output a whole bunch of information, including the transcript, and saved the text version of that transcript (no timestamps) to input.wav.txt.

How I figured out the 16khz MP3 conversion#

When I realized Whisper needed WAV and not MP3 I used my llm cmd command to figure out how to run that conversion:

1
llm cmd convert input.mp3 to .wav

It suggested this command, which I ran:

1
ffmpeg -i input.mp3 input.wav

But when I ran the resulting file through Whisper I got this error:

1
...
2
read_wav: WAV file '/tmp/input.wav' must be 16 kHz
3
error: failed to read WAV file '/tmp/input.wav'

So I ran llm cmd again:

1
llm cmd convert input.mp3 to .wav must be 16khz

And this time it gave me:

1
ffmpeg -i input.mp3 -ar 16000 input.wav

Which produced a file that worked in Whisper.

whisper-cpp has a bunch more options#

Here’s the full whisper-cpp --help output. I have not spent any time exploring these options beyond --output-txt:

1
usage: whisper-cpp [options] file0.wav file1.wav ...
2

3
options:
4
  -h,        --help              [default] show this help message and exit
5
  -t N,      --threads N         [4      ] number of threads to use during computation
6
  -p N,      --processors N      [1      ] number of processors to use during computation
7
  -ot N,     --offset-t N        [0      ] time offset in milliseconds
8
  -on N,     --offset-n N        [0      ] segment index offset
9
  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
10
  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
11
  -ml N,     --max-len N         [0      ] maximum segment length in characters
12
  -sow,      --split-on-word     [false  ] split on word rather than on token
13
  -bo N,     --best-of N         [5      ] number of best candidates to keep
14
  -bs N,     --beam-size N       [5      ] beam size for beam search
15
  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
16
  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
17
  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
18
  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
19
  -tr,       --translate         [false  ] translate from source language to english
20
  -di,       --diarize           [false  ] stereo audio diarization
21
  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
22
  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
23
  -otxt,     --output-txt        [false  ] output result in a text file
24
  -ovtt,     --output-vtt        [false  ] output result in a vtt file
25
  -osrt,     --output-srt        [false  ] output result in a srt file
26
  -olrc,     --output-lrc        [false  ] output result in a lrc file
27
  -owts,     --output-words      [false  ] output script for generating karaoke video
28
  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
29
  -ocsv,     --output-csv        [false  ] output result in a CSV file
30
  -oj,       --output-json       [false  ] output result in a JSON file
31
  -ojf,      --output-json-full  [false  ] include more information in the JSON file
32
  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
33
  -ps,       --print-special     [false  ] print special tokens
34
  -pc,       --print-colors      [false  ] print colors
35
  -pp,       --print-progress    [false  ] print progress
36
  -nt,       --no-timestamps     [false  ] do not print timestamps
37
  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
38
  -dl,       --detect-language   [false  ] exit after automatically detecting language
39
             --prompt PROMPT     [       ] initial prompt
40
  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
41
  -f FNAME,  --file FNAME        [       ] input WAV file path
42
  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
43
  -ls,       --log-score         [false  ] log best decoder scores of tokens
44
  -ng,       --no-gpu            [false  ] disable GPU