lfm2-5-audio-jp-ss-jdsc-asr-concat-tts-v1

Fine-tuned LiquidAI/LFM2.5-Audio-1.5B-JP for Japanese dysarthric ASR on SS-JDSC.

Training

• Base model: LiquidAI/LFM2.5-Audio-1.5B-JP
• Run: asr_ft_concat_tts_v1
• Prompt: Perform ASR in Japanese.
• Data: SS-JDSC train manifest plus concat augmentation and 1,000 TTS synthetic samples
• Checkpoint source: Modal Volume ss-jdsc-asr

Evaluation

• Full test rows: 1363
• CER: 44.39%
• test_basic: 57.5% (468 rows)
• test_daily: 32.13% (448 rows)
• test_hard: 43.0% (447 rows)

Files

This repo contains the fine-tuned Liquid Audio checkpoint files copied from:

/vol/ss_jdsc/checkpoints/asr_ft_concat_tts_v1/final

Evaluation artifacts are included under eval/asr_ft_concat_tts_v1/ when available.