Indic Whisper Ggml

--- license: apache-2.0 language: - kn - hi - ta - te - gu library_name: whisper.cpp tags: - whisper - whisper.cpp - ggml - speech-to-text - indic --- # Indic Whisper — ggml builds for whisper.cpp `ggml` + `q5_1` quantized builds of Indian-language Whisper fine-tunes, for on-device speech-to-text via [whisper.cpp](https://github.com/ggerganov/whisper.cpp). Used by the Ukta in-store feedback kiosk for accurate regional-language transcription. ## Files File naming: `ggml- -small.bin` — q5_1 quantized, ~181 MiB each. | Language | Code | File | |----------|------|------| | Hindi | hi | `ggml-hi-small.bin` | | Kannada | kn | `ggml-kn-small.bin` | | Tamil | ta | `ggml-ta-small.bin` | | Telugu | te | `ggml-te-small.bin` | | Gujarati | gu | `ggml-gu-small.bin` | These are **monolingual** — each model transcribes only its own language. Malayalam/Marathi/Odia/Punjabi/Bengali are not yet covered (no published `vasista22` small fine-tune); those languages fall back to a general model. ## Provenance & attribution - Fine-tuned source models: **[vasista22](https://huggingface.co/vasista22)** (`whisper- -{base,small}`), © Speech Lab, IIT Madras — **Apache 2.0**. - Base architecture/weights: **[OpenAI Whisper](https://github.com/openai/whisper)** — MIT. - Training corpora include AI4Bharat datasets (Shrutilipi, Vistaar) and Fleurs (CC-BY). - Conversion: `whisper.cpp/models/convert-h5-to-ggml.py` → f16 ggml, then `quantize ... q5_1`. This repository redistributes derivatives of the above under the **Apache License 2.0**; see `LICENSE`. No change was made to the model weights other than format conversion and quantization.