JacobPercy/st_aeds_asr_whisper_medium_finetune

ST-AEDS ASR Whisper Medium Fine-tuned Model

ESPnet2 ASR model fine-tuned on ST-AEDS / OpenSLR SLR45.

Model

• Base model: OpenAI Whisper Medium
• Training: full fine-tuning
• Decode: beam size 10, no CTC, no LM
• Token type: whisper_multilingual
• Language: English

Results

split	WER	CER
dev	3.9	2.7
test	2.1	0.9

Data

• Dataset: ST-AEDS / OpenSLR SLR45
• Train: 3038 utterances
• Dev: 400 utterances
• Test: 400 utterances
• No normalized transcript overlap was found across train/dev/test.

Files

• valid.acc.ave_3best.pth: averaged ESPnet ASR model checkpoint
• config.yaml: ESPnet training config
• RESULTS.md: ESPnet evaluation output