Whisper.cpp Audio Transcription Works | Generated by AI
🔔 Need to build a product? Check out truthsoft.dev — billed by token, collaborate anytime, your long-term technical partner. Truthsoft, an AI-native company: model training & inference, AI consulting, AI apps & agents, general app dev. Collaborate anytime · Do worthwhile things · Play the long game
Smoke run of whisper.cpp itself — audio → text. Everything works.
1. Basic transcription (WAV)
./build/bin/whisper-cli -m models/ggml-tiny.en.bin -f samples/jfk.wav
[00:00:00.000 --> 00:00:07.960] And so my fellow Americans ask not what your country can do for you
[00:00:07.960 --> 00:00:10.760] ask what you can do for your country.
load time = 148.84 ms | encode = 890.24 ms | decode = 38.99 ms | total = 1951.03 ms
Correct transcript, ~1.95 s for 11 s of audio (tiny.en, 4 threads, Vulkan backend present but CPU path used).
2. MP3 input (miniaudio decode path)
./build/bin/whisper-cli -m models/ggml-tiny.en.bin -f samples/jfk.mp3 -nt
→ And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.
3. Output formats
-osrt -ovtt -oj -of /tmp/jfk produced valid .srt, .vtt, .json:
1
00:00:00,000 --> 00:00:07,960
And so my fellow Americans ask not what your country can do for you
2
00:00:07,960 --> 00:00:10,760
ask what you can do for your country.
4. Other flags / examples
| Run | Result |
|---|---|
-t 2 -bs 1 -np (greedy, 2 threads) |
same text, segments snapped to 8.0/11.0 s |
-ml 1 (word timestamps) |
OK |
whisper-vad-speech-segments -vm models/for-tests-silero-v6.2.0-ggml.bin -f samples/jfk.wav |
detected 4 speech segments (0.32–2.27 s, 3.27–4.41 s, 5.38–7.68 s, 8.16–10.62 s) |
Notes
- Only real model on disk is
models/ggml-tiny.en.bin(English-only), so no multilingual/translate test; thefor-tests-ggml-*.binfiles are empty stubs (they load but produce"", as intended for CI). - Only
samples/jfk.wav/jfk.mp3are available; noespeak/ffmpeg/soxinstalled to synthesize extra audio. - Working tree left clean (temp outputs written to
/tmp).
Verdict: audio→text works end to end (WAV + MP3 in, plain/SRT/VTT/JSON out, VAD segmentation OK).
🌍 We created WeChat and WhatsApp groups for blog readers — welcome aboard! Scan the QR codes to join and meet visitors from China, Singapore, USA and around the globe to talk technology, AI, and connect with like-minded people.
WeChat Group
WhatsApp Group