Inworld TTS
Emotional multilingual speech in 0.25s — claimed #1 on HF TTS Arena
Generates emotional multilingual speech in 0.25 seconds; claimed #1 on the HF TTS Arena (per the post — verify before writing).
Voice AI: lifelike TTS, accurate ASR, and real-time voice agents.
Key features
- Training code for Inworld TTS-1 and TTS-1-Max models
- Pre-train, fine-tune or RL-align arbitrary SpeechLM-based TTS
- Modeling with SpeechLM and 1D audio codecs
- Distributed training with DDP, DeepSpeed and FSDP
Good for
- ML teams training custom speech language models
- Researchers fine-tuning TTS on multi-GPU clusters
Stack & tags
PythonPyTorchCUDADeepSpeedvLLMtts-trainingspeechlmdistributed-trainingrlhfaudio-codecopen-source
Quick start
make installFrom the project README on GitHub.