Model reference · open weights

distill-whisper-th-small

Audio biodatlab Speech→text 1 build Open weights 1k dl/mo

distill-whisper-th-small is an open-weight audio or speech model from biodatlab. distill-whisper-th-small (FP16) weighs 412 MB; the smallest configuration that runs it is RTX 3060 12 GB.

Hugging Face ↗

What it is

Released by	biodatlab
Type	Audio & music
Task	Speech→text
Parameters (lead)	206M
Runs with	transformers
Released	2024-01-16
Popularity	1k downloads / month
Weights	412 MB (distill-whisper-th-small (FP16), file size)
Licence	Open weights

What it runs on

Memory and cards for distill-whisper-th-small (FP16)

Weights 412 MB (file size) · overhead about 1.6 GB.

Card	One stream	Counted memory
RTX 3060 12 GB	fits	11.6 GB
RTX 4060 Ti 16 GB	fits	15.4 GB
RTX 3090 24 GB	fits	23.4 GB
RTX 4090 24 GB	fits	23.4 GB
RTX 5090 32 GB	fits	31.0 GB
L40S 48 GB	fits	44.0 GB
A100 80 GB	fits	78.2 GB
H100 80 GB	fits	78.1 GB
RTX PRO 6000 Blackwell 96 GB	fits	93.8 GB
DGX Spark (GB10) 128 GB unified	fits	107 GB
H200 141 GB	fits	138 GB
B200 180 GB	fits	176 GB

Estimates, not measurements: the weights are the build's file size. A speech model's decoder keeps a small cache for every stream it transcribes, so memory grows with the streams and beams at once. Counted memory is 92 % of what CUDA reports for the card.

From the model card

What biodatlab says about distill-whisper-th-small

Distilled Small Whisper ASR Model for Thai

Model Description

This is a distilled Automatic Speech Recognition (ASR) model, based on the Whisper architecture. It has been specifically tailored for Thai language speech recognition. The model features 4 decoder layers (vs 12 in teacher model) and has been distilled from a larger teacher model, focusing on enhancing performance and efficiency.

Distillation Details

Teacher Model: Small Whisper ASR model
Datasets Used for Distillation:
- Common Voice v13
- Gowajee
- Thai Elderly Speech Corpus
- Custom Scraped Data
- Thai-Central Dialect from SLSCU Thai Dialect Corpus

Model Performance

DeepCut Tokenized WER on Common Voice 13 Test Set:
- Distilled Model: 11.23%
- Teacher Model: 13.14%

This shows an improvement in Word Error Rate (WER), indicating enhanced accuracy in speech recognition tasks for the Thai language.

Read the full model card

Intended Use

This model is intended for use in applications requiring Thai language speech recognition.

Limitations

The model is specifically trained for the Thai language and may not perform well with other languages.
Performance might vary across different Thai dialects and accents.
As with any ASR system, background noise and speech clarity can impact recognition accuracy.

Acknowledgments

This model was developed using resources and datasets provided by the speech and language technology community. Special thanks to the teams behind Common Voice, Gowajee, SLSCU, and the Thai Elderly Speech Corpus for their valuable datasets.

Framework versions

Transformers 4.35.2
Pytorch 2.1.2
Datasets 2.16.1
Tokenizers 0.15.0

Contributors

Citation

Cite using Bibtex:

@misc {thonburian_whisper_med,
    author       = { Atirut Boribalburephan, Zaw Htet Aung, Knot Pipatsrisawat, Titipat Achakulvisut },
    title        = { Thonburian Whisper: A fine-tuned Whisper model for Thai automatic speech recognition },
    year         = 2022,
    url          = { https://huggingface.co/biodatlab/distil-whisper-th-small },
    doi          = { 10.57967/hf/0226 },
    publisher    = { Hugging Face }
}

Quoted from the model card on Hugging Face — the full card is behind the Hugging Face link above.