Nava for whisper.cpp
Nava, a Persian speech-to-text model, converted for whisper.cpp. Results, training details and limitations are on the main model page.
| File | Size | Notes |
|---|---|---|
ggml-nava.bin |
1.51 GB | 16-bit float |
ggml-nava-q8_0.bin |
0.81 GB | 8-bit quantized |
ggml-nava-q5_0.bin |
0.53 GB | 5-bit quantized |
huggingface-cli download Nimruz/Nava-ggml ggml-nava-q5_0.bin --local-dir .
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav
whisper-cli -m ggml-nava-q5_0.bin -l fa -nt -f audio.wav
Things to know
- Always pass
-l fa -nt. Nava was fine-tuned with the Persian prompt and without timestamp tokens. - Audio longer than 30 seconds needs splitting at quiet points into pieces of at most 30 seconds before it is passed in. The main repo has a Python script that does this.
- Output has no punctuation and spells numbers out. The main repo has a formatter that turns spoken numbers, dates and times into digits.
The files were built and tested with whisper.cpp v1.7.4. Each gave the same text as the original model on 9 of 10 FLEURS test clips, with a difference of a word or two on the tenth.