sna-whisper-asr

Shona–English Whisper (manassehzw/sna-whisper-asr)

Full merged weights (LoRA fused into unsloth/whisper-large-v3-turbo). Load with Transformers only — no PEFT or Unsloth at inference.

Training summary

• Base model: unsloth/whisper-large-v3-turbo
• Fine-tuning: LoRA (Unsloth + Transformers), merged for distribution
• Labels: Normalized transcripts (lowercase, punctuation stripped)
• Approx. eval WER (held-out subset, training setup): ~39–40% (300-example held-out eval, normalized references)

Load locally

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

MODEL_ID = "manassehzw/sna-whisper-asr"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
)
model.eval()

asr = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch.float16,
    device=0 if torch.cuda.is_available() else -1,
)

out = asr("your.wav", generate_kwargs={"task": "transcribe"})
print(out["text"])

Clone weights

git lfs install
git clone https://huggingface.co/manassehzw/sna-whisper-asr