CSM-1B — GRPO Naturalness Optimization

Sesame CSM-1B optimized with Group Relative Policy Optimization (GRPO) using UTMOS as a naturalness reward signal.

Why GRPO for TTS?

Standard supervised fine-tuning minimizes reconstruction loss, which doesn't always correlate with perceived naturalness. GRPO instead:

1. Generates multiple speech candidates per text input
2. Scores each with UTMOS (a trained naturalness predictor)
3. Uses relative rankings to update the policy toward higher-quality outputs

This is inspired by GLM-4-Voice which showed RL-based optimization improves TTS quality metrics beyond supervised fine-tuning.

Training

• Base: sesame/csm-1b (1B params)
• Method: GRPO with UTMOS reward (v3 — third iteration with refined hyperparams)
• Steps: 1,500 (best checkpoint by reward)
• Hardware: NVIDIA A10G 24GB

Files

Part of Project Maya

• Maya CSM-1B Pipeline
• Maya Orpheus Pipeline
• Decoder-only SFT
• Combined naturalness SFT