Modelpublic

mudler/magpie-tts.cpp-gguf

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

1likes2.6kdownloads

Model Card

magpie-tts.cpp GGUF

Brought to you by the [LocalAI](https://github.com/mudler/LocalAI) team, the folks behind LocalAI, the open-source AI engine that runs any model (LLMs, vision, voice, image, video) on any hardware, no GPU required.

Self-contained GGUF builds of NVIDIA's Magpie TTS Multilingual 357M for magpie-tts.cpp, a from-scratch C++17/ggml inference engine. Each file bundles the TTS model, the NanoCodec decoder, the tokenizer and the G2P dictionaries: one file, no Python, PyTorch, NeMo, or CUDA toolkit at inference.

5 voices (Aria, Jason, John, Leo, Sofia), 22.05 kHz mono, 9+ languages (en, es, de, fr, it, pt-BR, hi, vi, ko, ar variants; zh/ja not yet supported by the C++ tokenizer).

Files

File	Size	Notes
magpie-tts-multilingual-357m-f32.gguf	1294 MB	lossless, parity reference (teacher-forced replay max abs diff 3.6e-5 vs NeMo)
magpie-tts-multilingual-357m-f16.gguf	784 MB	ASR round-trip exact
magpie-tts-multilingual-357m-q8_0.gguf	624 MB	recommended; ASR round-trip exact, ~1.6x faster decode than f32
magpie-tts-multilingual-357m-q6_k.gguf	584 MB	ASR round-trip exact
magpie-tts-multilingual-357m-q5_k.gguf	562 MB	ASR round-trip exact on the smoke set; larger logit drift, expect degradation on hard material
magpie-tts-multilingual-357m-q4_k.gguf	541 MB	smallest; same caveat as q5_k

Quantization is selective (only matmul weights; codec and embeddings stay f32). Full drift numbers and methodology: docs/quantization.md.

Usage

bash

git clone --recursive https://github.com/mudler/magpie-tts.cpp
cd magpie-tts.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j

./build/examples/cli/magpie-cli say \
  --model magpie-tts-multilingual-357m-q8_0.gguf \
  --text "Hello world!" --lang en --speaker Aria --output hello.wav

Performance

About 66x faster than the NeMo reference pipeline on the same CPU (Ryzen 9 9950X3D: 12.1 s vs 805.7 s for ~4 s of speech, f32). Honest methodology and caveats: benchmarks/BENCHMARK.md.

License

Inference code: MIT. Model weights (these GGUFs): NVIDIA Open Model License. Model and codec by NVIDIA (NeMo team).

Built by the LocalAI team. If you want to run text to speech (and LLMs, vision, voice, image, and video models) locally on any hardware with an OpenAI-compatible API, give LocalAI a star.