metadata
pipeline_tag: automatic-speech-recognition
language: mri
license: apache-2.0
tags:
- trimmed
- whisper
library_name: transformers
base_model: openai/whisper-large-v3-turbo
base_model_relation: quantized
whisper-large-v3-turbo-mri-32768
This model is a vocabulary-pruned version of openai/whisper-large-v3-turbo optimised for Maori via the trimming method.
Only the decoder token-embedding table and lm_head (proj_out) were pruned. The encoder is fully unchanged, so audio feature extraction is identical to the original model.
Vocabulary reduction
| Component | Original | Pruned |
|---|---|---|
| vocab_size | 50257 | 31159 |
| embed_tokens | 66.4M | 41.9M |
| proj_out | 0.0M | 0.0M |
| Total | 808.9M | 784.4M |
Usage
from transformers import WhisperForConditionalGeneration, AutoTokenizer, pipeline
model = WhisperForConditionalGeneration.from_pretrained("alphaedge-ai/whisper-large-v3-turbo-mri-32768")
tokenizer = AutoTokenizer.from_pretrained("alphaedge-ai/whisper-large-v3-turbo-mri-32768")
asr = pipeline("automatic-speech-recognition", model=model, tokenizer=tokenizer)
result = asr("audio.wav", generate_kwargs={"language": "mri"})
print(result["text"])
⚠️ Limitations
Tokens not used in the selected language(s) were removed. The model may produce incorrect output for other languages.