agentory/MOSS-TTS-Realtime-GGUF

MOSS-TTS-Realtime — GGUF

OpenMOSS's MOSS-TTS-Realtime: Chinese and English TTS.

Two files make one model, as with llama.cpp's multimodal projectors: the model file is the text-to-speech LM backbone, run by llama.cpp; the codec file holds the audio side (decoder / vocoder, speaker encoders and the adaptor between the two), run by codec.cpp. The runtimes below wire the two graphs together, on CPU or GPU (CUDA, Vulkan, Metal), fully offline.

Files

Every run needs one model file + one codec file from this repo.

	--backbone	--model (codec)	Total
Recommended	moss-tts-realtime-bf16.gguf	codec-q8_0.gguf	5.9 GB
Smaller	moss-tts-realtime-q8_0.gguf	codec-q4_k_m.gguf	3.4 GB
Best quality	moss-tts-realtime-bf16.gguf	codec-f16.gguf	7.6 GB

Use the bf16 model file where you can; lower-precision model files lose intelligibility first.

Other files

File	Size
codec-f32.gguf	7.7 GB
codec-q5_k_m.gguf	1.8 GB
moss-tts-realtime-f16.gguf	3.4 GB
moss-tts-realtime-f32.gguf	6.9 GB
moss-tts-realtime-q2_k.gguf	778 MB
moss-tts-realtime-q3_k_l.gguf	1.0 GB
moss-tts-realtime-q3_k_m.gguf	940 MB
moss-tts-realtime-q3_k_s.gguf	867 MB
moss-tts-realtime-q4_0.gguf	1.1 GB
moss-tts-realtime-q4_1.gguf	1.1 GB
moss-tts-realtime-q4_k_m.gguf	1.1 GB
moss-tts-realtime-q4_k_s.gguf	1.1 GB
moss-tts-realtime-q5_0.gguf	1.2 GB
moss-tts-realtime-q5_1.gguf	1.3 GB
moss-tts-realtime-q5_k_m.gguf	1.3 GB
moss-tts-realtime-q5_k_s.gguf	1.2 GB
moss-tts-realtime-q6_k.gguf	1.4 GB

Command line (codec.cpp)

With codec.cpp's tts-cli (build it; it bundles llama.cpp), then:

hf download BricksDisplay/MOSS-TTS-Realtime-GGUF moss-tts-realtime-bf16.gguf codec-q8_0.gguf --local-dir MOSS-TTS-Realtime

./build/tts-cli synthesize --backbone MOSS-TTS-Realtime/moss-tts-realtime-bf16.gguf --model MOSS-TTS-Realtime/codec-q8_0.gguf \
    --temp 0.8 --top-p 0.6 --top-k 30 --rep-penalty 1.1 \
    --text "你好，欢迎使用语音合成。" --output out.wav

All tts-cli options (voices, voice cloning, sampling, GPU) are described in the codec.cpp README.

Python

The codec_cpp package runs the same pipeline (it builds codec.cpp with its bundled llama.cpp, so it needs CMake and a C++ compiler):

pip install "codec-cpp @ git+https://github.com/mybigday/codec.cpp" huggingface_hub soundfile

import codec_cpp, soundfile
from huggingface_hub import hf_hub_download

tts = codec_cpp.TTS(hf_hub_download("BricksDisplay/MOSS-TTS-Realtime-GGUF", "moss-tts-realtime-bf16.gguf"),
                    hf_hub_download("BricksDisplay/MOSS-TTS-Realtime-GGUF", "codec-q8_0.gguf"))
audio, sr = tts.synthesize("你好，欢迎使用语音合成。",
                         temperature=0.8, top_p=0.6, top_k=30, repetition_penalty=1.1)
soundfile.write("out.wav", audio, sr)

tts.voices() lists the builtin voices; ref_audio= clones a voice where the model supports it; tts.stream(text) yields audio as it is generated. More in the Python docs.

React Native (llama.rn)

llama.rn loads the same two files on iOS and Android:

const ctx = await initLlama({ model: 'moss-tts-realtime-bf16.gguf', embedding: true })
await ctx.initVocoder({ path: 'codec-q8_0.gguf' })
const { prompt, grammar, embedding, flow } = await ctx.getFormattedAudioCompletion({
  prompt: 'Hello world',
})
const result = await ctx.completion({ prompt, grammar, embedding })
const pcm = flow === 'continuous_embd'
  ? await ctx.decodeAudioEmbeddings(result.embeddings, result.embedding_dim)
  : await ctx.decodeAudioTokens(result.audio_tokens)

Voice cloning, saved voices and per-model notes: llama.rn TTS docs.

Notes

The sampling flags in the example are the upstream settings; keep them.

Sampling

The model files carry the recommended sampling as llama.cpp's general.sampling.* metadata, the settings of the original streaming inference code; tts-cli, tts-server, llama.rn and any llama.cpp host apply them unless you set your own.

Setting	Value	tts-cli flag
temperature	0.8	--temp 0.8
top-k	30	--top-k 30
top-p	0.6	--top-p 0.6
min-p	0 (off)	--min-p 0.0
repetition penalty	1.1	--repeat-penalty 1.1
repetition window	50	--repeat-last-n 50
sampler order	penalties → temperature → top-k → top-p	--samplers "penalties;temperature;top_k;top_p"

Each of the 16 codebooks keeps its own repetition window.

License

apache-2.0, the same as the original model.