lfm2-5-audio-jp-ss-jdsc-asr-best-v1

Fine-tuned LiquidAI/LFM2.5-Audio-1.5B-JP for Japanese dysarthric ASR on the single-speaker SS-JDSC corpus. This is the project-best checkpoint (asr_ft_joint_warmstart_v1), trained with a two-stage warm-start joint recipe.

Output format (IMPORTANT)

The model emits a joint reading ⟶ kanji sequence: it first produces the spoken katakana reading (acoustic grounding), then the kanji-kana orthography after the ⟶ (U+27F6) marker. For a normal kanji transcript, split on ⟶ and keep the part after it:

text = hypothesis.split("⟶")[-1].strip()

Example: ワタシワシューニヨンカイ… ⟶ 私は週に四回…

Training recipe (two-stage, single model)

1. Stage 1 — katakana grounding: FT the base audio model on katakana pronunciation targets (UniDic pron, via fugashi), so audio→reading is strong and homophone-spelling hallucination is removed.
2. Stage 2 — warm-start joint: initialise from the Stage-1 checkpoint and FT on reading ⟶ kanji joint targets. The decoder learns the reading→kanji conversion in-model (no separate LLM), and the carried-over acoustic grounding absorbs the conversion error.

• Base model: LiquidAI/LFM2.5-Audio-1.5B-JP
• Run: asr_ft_joint_warmstart_v1 (warm-started from asr_ft_kana_v1)
• Prompt: Perform ASR in Japanese.
• 3000 steps, context length 512, full fine-tune (bf16), Modal L40S
• Checkpoint source: Modal Volume ss-jdsc-asr

Evaluation (kanji CER, full SS-JDSC test set)

Scored on the kanji part only (split on ⟶) vs. the original kanji reference.

Subset	Direct kanji FT	This model (warm-start joint)
total	44.39%	41.85%
basic	57.50%	51.23%
daily	32.13%	32.49%
hard	43.00%	41.99%

Project-best total CER (41.85%, −2.54 pt vs direct kanji FT), with the largest gain on the hallucination-prone basic subset (−6.27 pt).

Eval summary at upload time — rows: 1363, CER: 41.85%

• test_basic: 51.23% (468 rows)
• test_daily: 32.49% (448 rows)
• test_hard: 41.99% (447 rows)

Files

Fine-tuned Liquid Audio checkpoint files, copied from:

/vol/ss_jdsc/checkpoints/asr_ft_joint_warmstart_v1/final

Evaluation artifacts are included under eval/asr_ft_joint_warmstart_v1/ when available.