gripened/auderly-whisper-hinglish-ggml

Auderly — Whisper Hinglish GGML (q5_0)

whisper.cpp GGML (q5_0, ~1 GB) conversion of Trelis/whisper-hinglish-preview — a Whisper-large-v3 finetune that transcribes English+Hindi code-switched speech with English in Latin and Hindi in Devanagari in the same utterance. Packaged for on-device use (whisper.rn / whisper.cpp).

Provenance & credit

All model credit to Trelis/whisper-hinglish-preview (Whisper-large-v3 + <|mixedcode|> token; itself finetuned from ARTPARK-IISc Vaani-Hindi). This repo only re-packages it: HF safetensors → GGML (convert-h5-to-ggml.py) → quantized q5_0.

whisper.cpp patch (REQUIRED)

Vocab is 51867 (large-v3's 51866 + one <|mixedcode|> token at id 51866). Stock whisper.cpp mis-maps timestamps for n_vocab 51867. One-line fix so it keeps the v3 layout:

int num_languages() const {
    const int extra = (n_vocab == 51867) ? 2 : (is_multilingual() ? 1 : 0);
    return n_vocab - 51765 - extra;
}

Then it code-switches for any language token (whisper-cli -l hi works; no <|mixedcode|> injection needed).

License

Derivative of Trelis/whisper-hinglish-preview — see that repo's license. Credit: Trelis Research + the Vaani project (ARTPARK @ IISc).