Bengali Whisper Medium — MLX 8-bit
This model converts Bengali speech to text on Apple Silicon Macs. It uses the weights from BengaliAI regional Whisper Medium. This release changes the model format and weight precision. It includes no additional training.
| Property | Value |
|---|---|
| Language | Bengali |
| Model format | MLX |
| Weight precision | 8-bit affine quantization; group size 64 |
| Weight file size | 814 MB |
| Tested hardware | Apple M4 Pro; 20 GPU cores; 48 GB memory |
| Tested software | Python 3.13; MLX 0.32.1; MLX Audio 0.5.3 |
| License | Apache-2.0 |
Quantization reduces weight precision to decrease file size. Some tensors retain floating-point precision. The model requires more memory than its weight file size.
Installation
-
Create a Python environment.
python3 -m venv .venv -
Activate the environment.
source .venv/bin/activate -
Install the download tool.
pip install huggingface-hub -
Download the model.
hf download tuxboy/bengali-whisper-medium-mlx-8bit --local-dir bengali-whisper-medium-mlx-8bit -
Install the required packages.
pip install -r bengali-whisper-medium-mlx-8bit/requirements.txt
Transcription
Use a WAV or FLAC file.
Replace recording.wav with the path to your audio file.
python bengali-whisper-medium-mlx-8bit/transcribe.py recording.wav
The script converts audio to mono at 16 kHz. It divides audio at pauses into sections of 26 seconds or less. It processes each section separately. The maximum output is 440 tokens per section.
The script includes whisper_compat.py. This file corrects a token lookup error in MLX Audio 0.5.3.
Without this correction, the decoder can fail to stop. The correction does not change installed packages.
Speed test
Each test used the same 300-second recording and 13 audio sections. Each test ran once after a warm-up run.
| Runtime | Processing time | Peak process memory (RSS) |
|---|---|---|
| PyTorch/MPS fp16 | 43.78 s | 5.13 GB |
| MLX fp16 | 24.79 s | 2.12 GB |
| MLX 8-bit | 20.76 s | 1.41 GB |
Times include feature extraction and decoding. Times exclude model loading, file access, and audio conversion. These results apply to this test. They do not measure transcription accuracy.
Limits
- The model can produce incorrect words for silence or noise. The script disables the no-speech filter.
- Pause detection uses audio energy. Continuous speech can cause cuts within words.
- The 8-bit and fp16 outputs matched on 24 short OpenSLR53 clips. This small test does not establish general accuracy.
- Tests do not establish accuracy for each Bengali dialect.
- Android and Raspberry Pi operation have not been tested.
Source and license
BengaliAI trained the source model on Ben-10. See the source model card for the research citation.
Source revision: 76cce9a874e8752d561a7aff28f01383df270db4.
See LICENSE for the Apache-2.0 license. See export-provenance.json for conversion data. See SHA256SUMS.json for file checksums.