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

Fine-tuned LiquidAI/LFM2.5-Audio-1.5B-JP for single-speaker Japanese TTS on SS-JDSC.

Training

• Base model: LiquidAI/LFM2.5-Audio-1.5B-JP
• Run: v1
• Prompt: Perform TTS in japanese.
• Data: JDSC-Project/SS-JDSC train_13.0h single-speaker corpus
• Context length: 384
• Training: 3000 steps, batch size 32 (~9.5 epochs)
• Checkpoint source: Modal Volume ssjdsc-tts:/out/v1/final

Notes

This checkpoint is the TTS generator used to create synthetic SS-JDSC augmentation audio for ASR experiments. The basic and hard variants in downstream experiments are produced by changing the input text generation, not by using separate TTS checkpoints.

Files

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

/vol/out/v1/final