Zizi models
Speech recognition models used by Zizi, an on-device voice recorder and transcriber (Android and iOS). Every file is a quantized model in ggml format for whisper.cpp. The app downloads them directly from this repo, for example:
https://huggingface.co/byteview-co/zizi-models/resolve/main/ggml-small-q5_1.bin
Files
| File | Model | SHA-256 |
|---|---|---|
ggml-phowhisper-small-q5_1.bin |
PhoWhisper small (Vietnamese), q5_1 | c93fb04f6f586de4eb9b82501f371d4b0bdd033e266264a5ab4f787dea25343e |
ggml-phowhisper-medium-q5_0.bin |
PhoWhisper medium (Vietnamese), q5_0 | e2cd4ceda28646fc74e48f171fad8ebd75e0ab06817d70349bc42ae2335bfc27 |
ggml-small-q5_1.bin |
Whisper small (multilingual), q5_1 | ae85e4a935d7a567bd102fe55afc16bb595bdb618e11b2fc7591bc08120411bb |
ggml-large-v3-turbo-q5_0.bin |
Whisper large-v3-turbo (multilingual), q5_0 | 394221709cd5ad1f40c46e6031ca61bce88931e6e088c188294c6d5a55ffa7e2 |
Sources and licenses
These files are redistributed under the licenses of the original models:
- Whisper small and Whisper large-v3-turbo: by OpenAI, MIT License. The ggml conversions are unchanged copies of the files published in ggerganov/whisper.cpp (MIT License).
- PhoWhisper small and PhoWhisper medium: by VinAI Research,
BSD-3-Clause License. They were converted to ggml and quantized (q5_1 for small, q5_0 for medium)
with
tools/convert_phowhisper.shfrom the Zizi app source. PhoWhisper outputs lower-case text without punctuation and has no timestamp tokens; Zizi decodes it per voice-activity segment withno_timestamps.
If you use PhoWhisper, please cite:
@inproceedings{PhoWhisper,
title = {{PhoWhisper: Automatic Speech Recognition for Vietnamese}},
author = {Thanh-Thien Le and Linh The Nguyen and Dat Quoc Nguyen},
booktitle = {Proceedings of the ICLR 2024 Tiny Papers track},
year = {2024}
}
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support