Ming-3-TTS-0.5B
An open-source 0.5B-parameter text-to-speech model released by Bailing that supports both zero-shot voice cloning and natural-language voice design. It outputs 44.1 kHz audio, covers multiple languages and dialects, and can be deployed locally on devices with limited GPU memory.
Timeline (last 90 days)
Ming-3-TTS-0.5B is a TTS model with only 0.5B parameters that supports both zero-shot voice cloning and one-sentence voice design.
Ming-3-TTS-0.5B's audio output specification is 44.1kHz.
Ming-3-TTS-0.5B supports 7 languages as well as dialects such as Sichuanese and Cantonese.
After Q8 quantization, Ming-3-TTS-0.5B requires about 2GB of VRAM to run.
Ming-3-TTS-0.5B's architecture comprises 24 layers, with the 0.5B backbone using GTOA to output features frame by frame, then an 8-layer DiT performing flow-matching denoising in a 64-dimensional shallow space, and finally the Odo VAE restoring 44.1kHz waveforms.
The cloning part of Ming-3-TTS-0.5B uses CamPlus to extract 192-dimensional speaker vectors.
In official test data, Ming-3-TTS-0.5B achieves a zero-shot Chinese word error rate of 0.8% and an English one of 1.58%.
In official test data, Ming-3-TTS-0.5B's voice design score is 72.29%.
The 0.5B parameter capacity limits Ming-3-TTS-0.5B's performance on complex emotions, long-text stability and extreme timbres, so its ceiling is lower than that of large models.
The 8-bit quantized version of Ming-3-TTS-0.5B occasionally has background noise at the tail end.
All Facts (10)
Ming-3-TTS-0.5B is a TTS model with only 0.5B parameters that supports both zero-shot voice cloning and one-sentence voice design.
50%UnverifiedMing-3-TTS-0.5B's audio output specification is 44.1kHz.
50%UnverifiedMing-3-TTS-0.5B supports 7 languages as well as dialects such as Sichuanese and Cantonese.
50%UnverifiedAfter Q8 quantization, Ming-3-TTS-0.5B requires about 2GB of VRAM to run.
50%UnverifiedMing-3-TTS-0.5B's architecture comprises 24 layers, with the 0.5B backbone using GTOA to output features frame by frame, then an 8-layer DiT performing flow-matching denoising in a 64-dimensional shallow space, and finally the Odo VAE restoring 44.1kHz waveforms.
50%UnverifiedThe cloning part of Ming-3-TTS-0.5B uses CamPlus to extract 192-dimensional speaker vectors.
50%UnverifiedIn official test data, Ming-3-TTS-0.5B achieves a zero-shot Chinese word error rate of 0.8% and an English one of 1.58%.
50%UnverifiedIn official test data, Ming-3-TTS-0.5B's voice design score is 72.29%.
50%UnverifiedThe 0.5B parameter capacity limits Ming-3-TTS-0.5B's performance on complex emotions, long-text stability and extreme timbres, so its ceiling is lower than that of large models.
50%UnverifiedThe 8-bit quantized version of Ming-3-TTS-0.5B occasionally has background noise at the tail end.
50%