audio-cpp/Ming-3.0-TTS-0.5B-GGUF

Ming 3.0 TTS 0.5B GGUF

Star or watch audio.cpp on GitHub to follow model integration, merged code changes and inference improvements.

GGUF packages of Ming 3.0 TTS 0.5B for audio.cpp. Each file contains the Qwen backbone, flow-matching head, audio VAE, CAMPPlus speaker encoder, configuration, tokenizer and model spec. Output is mono at 44.1 kHz. No separate codec or Python runtime is required.

audio.cpp family: ming3_tts.

Upstream implementation: inclusionAI/Ming-3.0-tts, pinned to revision e999c39fa676722aaa9ea29b5aa19a90525b271f for the Python comparison.

Package	Precision
ming-3.0-tts-0.5b-bf16.gguf	Original checkpoint precision
ming-3.0-tts-0.5b-q8_0.gguf	Q8_0; BF16 stop head and F32 speaker encoder retained

Performance

Measured on an RTX 5090 with eight CPU threads and an audio.cpp Debug C++17 build. Python uses the pinned official BF16 implementation above. Both use the same text, voice-design instruction and seed. Warm RTF covers a 154-character request, its repeat and a 40-character request; peak memory covers the complete session, including warmup and a 571-character request. The long request is excluded from the speed comparison because text chunking differs between implementations. Lower RTF is better.

Implementation	Warm RTF	Speed vs Python	Peak VRAM MiB	Peak CPU RSS MiB
Official Python, BF16	0.651-0.675	1x	3,660	2,104
audio.cpp, BF16	0.173-0.179	3.8x	2,768	1,036
audio.cpp, Q8_0	0.167-0.175	3.9x	1,924	978

Speed ratios use mean warm RTF. These are synthesis-speed measurements, not claims of identical sampled audio.

Usage

audiocpp_cli --family ming3_tts --task tts --backend cuda \
  --model ming-3.0-tts-0.5b-bf16.gguf \
  --request-option language=en \
  --text "Good evening. The harbor is quiet tonight." --out speech.wav --log

audiocpp_cli --family ming3_tts --task clon --backend cuda \
  --model ming-3.0-tts-0.5b-bf16.gguf --voice-ref reference.wav \
  --request-option language=en \
  --text "Please leave the lamp beside the window." --out clone.wav --log

audiocpp_cli --family ming3_tts --task vdes --backend cuda \
  --model ming-3.0-tts-0.5b-bf16.gguf \
  --request-option language=en \
  --request-option instruct="A calm adult man with a clear, warm voice." \
  --text "Good evening. The harbor is quiet tonight." --out designed.wav --log

Cloning without a transcript uses speaker conditioning. To include the reference's acoustic prefix, also supply --request-option reference_text="Words spoken in the reference.".

The model supports structured voice, emotion, dialect and speaking-attribute controls. speed is a categorical conditioning control, not waveform resampling. See the model guide for option values and long-form behavior.

The model weights retain their upstream MIT license; see LICENSE.