Indian-language Whisper models (ggml)

Fine-tuned Whisper speech-recognition models for 11 Indian languages and Hinglish, converted to whisper.cpp ggml format and quantized by AJtech. They are built and tested for whisper.cpp v1.9.5 and are used for on-device captioning in AJtech's apps.

These are modified versions of third-party models. Each file was converted from PyTorch/safetensors to ggml and quantized (q5_1 or q5_0). The weights were not otherwise changed. All credit for training goes to the original authors listed below.

Files

File Language (whisper code) Base Quant Size (bytes) SHA-256 License
ggml-hindi-small-q5_1.bin Hindi (hi) small q5_1 190,085,487 6813fed7ffa6c3fa14490c1f1788d2d8b6e3b7badf59a2f75fb4c5c21cf00f3f Apache-2.0
ggml-gujarati-small-q5_1.bin Gujarati (gu) small q5_1 190,085,487 e756347b8fb573b88a5625ad98dac9075746549fdf407260d321d077a73b094a Apache-2.0
ggml-hinglish-base-q5_1.bin Hindi speech โ†’ Roman Hinglish (en) base q5_1 59,707,625 30ba7b988bb0f7b3e3abc4b8aca1a5028cf56060d2142233d240c10f3d5b070a Apache-2.0
ggml-hinglish-turbo-q5_0.bin Hindi speech โ†’ Roman Hinglish (en) large-v3-turbo q5_0 574,041,195 9d877151b15cec1feb9110cfbc0a3162cf377bcc0ab1935174226f461cf60f13 Apache-2.0
ggml-tamil-small-q5_1.bin Tamil (ta) small q5_1 190,085,487 227e8a3b257de6e49a62c736708f89812573b6a39142f831cb9e7815a57dbfa2 Apache-2.0
ggml-telugu-small-q5_1.bin Telugu (te) small q5_1 190,085,487 47369abd7ee13b624606b762a860a42d7cbea8f320e3c4553954d1fea748d49e Apache-2.0
ggml-kannada-small-q5_1.bin Kannada (kn) small q5_1 190,085,487 55ed42b66165af8caa569dbf982bb528c4e1992338a9b8da71c64b6412cf6b4c Apache-2.0
ggml-bengali-small-q5_1.bin Bengali (bn) small q5_1 190,085,487 81b476996fc7fb0f14a14c4aa8fc657d4e677140ebaefeca2258c90601879e2c Apache-2.0
ggml-urdu-small-q5_1.bin Urdu (ur) small q5_1 190,085,487 2bc8261fde5198a5dbfecb8108d1954401aa894a1a240eca893499ba4684768c Apache-2.0
ggml-punjabi-medium-q5_0.bin Punjabi, Gurmukhi (pa) medium q5_0 539,212,467 bff57370d60a0a7d7941ff2c4b0cc68e954a1917239e81bfc4a49830c257800b MIT
ggml-malayalam-medium-q5_0.bin Malayalam (ml) medium q5_0 539,212,467 1a1ba89f8354ac0359ddc841c39b5cb51f2e246e02d718aa8c5707e98fe81394 MIT
ggml-marathi-medium-q5_0.bin Marathi (mr) medium q5_0 539,212,467 5e621293114abd3a2a07bc2e9b63828577d99b23b4973145ea34910627e2a53a MIT
ggml-odia-medium-q5_0.bin Odia (decode with hi) medium q5_0 539,212,467 02be2e37ebb78ea6e104bc8d4215c52f15e796b7cd05f7701bc8c70691e61966 MIT

Licenses

This repository has mixed licenses:

  • Apache-2.0 (see LICENSE): Hindi, Gujarati, Hinglish base, Hinglish turbo, Tamil, Telugu, Kannada, Bengali, Urdu.
  • MIT (see LICENSE-MIT): Punjabi, Malayalam, Marathi, Odia. Copyright (c) 2023 AI4Bhฤrat.
  • All files derive from OpenAI Whisper, MIT, Copyright (c) 2022 OpenAI (also in LICENSE-MIT).

Each file was modified from its source: format conversion to ggml and quantization.

Sources and credits

ggml file Source Revision converted Author
Hindi vasista22/whisper-hindi-small fb4a24afc20c42906deed31c3689e4be31da41fe Vasista Sai Lodagala
Gujarati vasista22/whisper-gujarati-small f08036e339f9d1a53b1e28f556e4b436d77f7762 Vasista Sai Lodagala
Tamil vasista22/whisper-tamil-small ac0d71cd70d4c66aec7126996cd9e56e97e698b7 Vasista Sai Lodagala
Telugu vasista22/whisper-telugu-small 717212f9c4c16a78a86ddcb98f93c7f73ef8a5cc Vasista Sai Lodagala
Kannada vasista22/whisper-kannada-small 6f8e7b770513d941e4713b20adde26af32b2970b Vasista Sai Lodagala
Hinglish base Oriserve/Whisper-Hindi2Hinglish-Swift cd0d4d013062d542569135460a79025337df1982 Oriserve
Hinglish turbo Oriserve/Whisper-Hindi2Hinglish-Apex f3214eed20b4e4d4144e739982d911f87b9cb223 Oriserve
Bengali Rakib/whisper-small-bn-crblp 55b78e2ee3f6446e924ca71486060ba188e4b255 Mohammed Rakib
Urdu khawajaaliarshad/whisper-small-urdu 163595c44c9e1def15c7804241a26174672c88f0 Ali Khawaja
Punjabi, Malayalam, Marathi, Odia AI4Bharat/vistaar (IndicWhisper) repo commit 21dce3406fddf0e18984165e735d23db38760dd3 AI4Bharat

The Hindi, Gujarati, Tamil, Telugu and Kannada models were trained by Vasista Sai Lodagala at Speech Lab, IIT Madras, with compute funded by the "Bhashini: National Language translation Mission" project of the Ministry of Electronics and Information Technology (MeitY), Government of India. IndicWhisper was built by AI4Bharat with support from the EkStep Foundation.

Usage notes (whisper.cpp)

  • Run the fine-tuned Indic models with no_timestamps = true. With timestamps on, they emit a junk first token. Take word timing from token_timestamps = true or DTW instead.
  • Feed them short chunks (โ‰ค 20 s, split at pauses), not fixed 30 s windows. They were trained on short utterances.
  • Native-script output can hit whisper.cpp's 220-token limit per 30 s window. Short chunks avoid this.
  • Decode the Odia model with language hi. Decode the Hinglish models with language en.
  • Verify the SHA-256 above after every download.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for alexthe4566/indic-whisper-ggml

Finetuned
(808)
this model