srinikvs/maata-telugu-small

Maata Telugu small (whisper.cpp ggml q8_0)

Telugu speech-to-text model for whisper.cpp: fine-tuned from openai/whisper-small, exported to ggml and quantized to q8_0.

• File: ggml-maata-small-q8_0.bin (264464624 bytes, sha256 57b5edfa39368038c5c86f27e7cdef026f0f9c9080cd9611a970374ef95f0e3c)
• Usage: whisper-cli -m ggml-maata-small-q8_0.bin -l te -f audio.wav (16 kHz mono WAV)

This repo contains only the model weights; no audio, clips, transcripts or training data. Access is gated: sign in to Hugging Face and share your contact details to download.

Credits and licenses

Fine-tuned from openai/whisper-small on the Telugu subset of IndicVoices by AI4Bharat.

• Base model: OpenAI Whisper, MIT License. The full copyright and license text is in NOTICE.
• Fine-tuning data: IndicVoices (Telugu train split) by AI4Bharat, IIT Madras, licensed CC-BY-4.0.
• Changes made: the Whisper small weights were changed by full fine-tuning on IndicVoices Telugu, then converted to whisper.cpp ggml format and quantized to q8_0.

Citation

@article{javed2024indicvoices,
  title   = {IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages},
  author  = {Javed, Tahir and Nawale, Janki Atul and George, Eldho Ittan and others},
  journal = {arXiv preprint arXiv:2403.01926},
  year    = {2024}
}