zortzi-tts-onnx

ONNX export of itzune/zortzi-tts — the Basque fine-tune of Audio8-TTS-Preview-0.6b. Ships INT4-quantized models (575 MB total) that run entirely on CPU with the Audio8 ONNX runtime. No GPU or PyTorch dependency.

Two voices: Maider (female) and Antton (male).

🏗️ Source code, export & quantization scripts: github.com/itzune/zortzi-tts

---

Files

├── slow_ar_int4.onnx (+ .data)         # Slow AR, INT4 (278 MB)
├── fast_ar_int4.onnx (+ .data)         # Fast AR, INT4 (34 MB)
├── slow_ar_fp16.onnx (+ .data)         # Slow AR, FP16 (1077 MB) — optional, higher fidelity
├── fast_ar_fp16.onnx (+ .data)         # Fast AR, FP16 (134 MB) — optional, higher fidelity
├── codec_decoder_fp16.onnx             # codes → audio (266 MB)
├── voices/
│   ├── maider/{codes.npy, meta.json}   # Maider reference voice codes
│   └── antton/{codes.npy, meta.json}   # Antton reference voice codes
├── runtime_manifest.json               # precision & architecture metadata
└── tokenizer/
    └── tokenizer.json

Precision comparison

Component	FP16	INT4	Notes
Slow AR	1077 MB	278 MB	121 MatMulNBits + 11 GatherBlockQuantized
Fast AR	134 MB	34 MB	21 MatMulNBits + 1 GatherBlockQuantized
Codec decoder	266 MB	266 MB	not quantized (Conv1d-based)
Inference total	1477 MB	575 MB	INT4 matches official Audio8 (572 MB)

Which to use? INT4 is the default — 2.5× smaller and ~28% faster on CPU. FP16 gives slightly better prosody fidelity. If you need FP16, download the *_fp16.onnx files and set default_precision: "fp16" in runtime_manifest.json. Prefer FP16 when fidelity matters; use INT4 when size or CPU speed is the priority.

---

Quick start (CPU — ONNX Runtime)

Install

# Clone the Audio8_TTS repo for the ONNX runtime (inference code)
git clone https://github.com/Audio8-AI/Audio8_TTS.git
cd Audio8_TTS/onnx_runtime

uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python onnxruntime soundfile numpy scipy

Synthesize

# Download this model:
# huggingface-cli download itzune/zortzi-tts-onnx --local-dir ./zortzi-tts-onnx

.venv/bin/python -m arktts_runtime.cli \
    --model-dir ./zortzi-tts-onnx \
    --voices-dir ./zortzi-tts-onnx/voices \
    --text "Kaixo mundua! Nire izena Maider da." \
    --voice maider \
    --output out.wav \
    --temperature 0.8 --top-p 0.95 \
    --max-new-tokens 512

Switch voices

# Antton
.venv/bin/python -m arktts_runtime.cli \
    --model-dir ./zortzi-tts-onnx \
    --voices-dir ./zortzi-tts-onnx/voices \
    --text "Kaixo mundua! Nire izena Antton da." \
    --voice antton \
    --output out.wav \
    --temperature 0.8 --top-p 0.95

Select precision

# INT4 (default — smaller + faster)
.venv/bin/python -m arktts_runtime.cli --model-dir .../zortzi-tts-onnx --precision int4 ...

# FP16 (higher fidelity — requires the *_fp16.onnx files)
.venv/bin/python -m arktts_runtime.cli --model-dir .../zortzi-tts-onnx --precision fp16 ...

The available precisions are declared in runtime_manifest.json (available_precisions). ONNX Runtime upcasts FP16 → FP32 internally for computation, so FP16 is a storage-only optimization on CPU.

Streaming

The Audio8 ONNX runtime includes a FastAPI streaming server:

cd /path/to/Audio8_TTS/onnx_runtime
.venv/bin/python -m arktts_runtime.service \
    --model-dir ./zortzi-tts-onnx \
    --voices-dir ./zortzi-tts-onnx/voices \
    --port 8024

• POST /api/tts/stream — streaming NDJSON with base64 PCM chunks (chunk_frames=12, ~557 ms of audio per chunk)
• POST /v1/audio/speech — OpenAI-compatible non-streaming
• GET /api/health — health check

CPU inference speed

On an 8-core CPU, a ~3s Basque utterance synthesizes in:

Precision	Time	RTF	Real-time?
INT4	~16s	5.3×	❌ no
FP16	~23s	7.7×	❌ no

CPU inference is not real-time. For real-time synthesis, use the PyTorch model on GPU (RTF ≈ 1.05× on an NVIDIA L40). CPU ONNX is intended for offline/batch generation where latency is acceptable.

---

How INT4 quantization works

Audio8's official INT4 models use standard ONNX Runtime operators — MatMulNBits (com.microsoft domain, bits=4, block_size=128) for linear weights and GatherBlockQuantized for embeddings. These are not proprietary. The quantization script (quantize_int4.py) runs two passes:

1. Linear weights — MatMul → MatMulNBits (121 nodes slow AR, 21 fast AR)
2. Embedding tables — Gather → GatherBlockQuantized (11 nodes slow AR, 1 fast AR)

This matches Audio8's official INT4 format exactly: 575 MB vs 572 MB official.

---

Limitations & known issues

• Numerals are not pronounced in Basque. Even with text normalization, the model pronounces numbers in a mix of languages. Spell out numbers manually.
• No built-in text normalization. Expand acronyms (e.g. "TTS" → "te te ese") upstream of the model.
• INT4 can degrade quality. Flatter prosody and less reliable question intonation compared to FP16. Prefer FP16 when fidelity matters.
• Two voices only. Maider and Antton.
• Sampling required. Always use --temperature 0.8 --top-p 0.95. Do not use greedy decoding (repetition loops).
• Not real-time on CPU. Use GPU PyTorch for real-time synthesis.

---

Model details

Architecture	DualAR TTS (Qwen2-style decoder + codec)
Parameters	601M (AR) + codec
Opset	17
Codebooks	10 × 4096 entries
Audio	44.1 kHz, ~21.5 frames/s
Max sequence length	2048
PyTorch source	itzune/zortzi-tts
Base model	Audio8/Audio8-TTS-Preview-0.6b (Apache-2.0)
License	Apache-2.0 (inherited from base)

---

Datasets, licenses & acknowledgements

Dataset	Role	License	Clips
Audio8-TTS-Preview-0.6b	Base model	Apache-2.0	—
Mozilla Common Voice 26.0 — Basque ¹	Phase 1: phonotactics	CC0 1.0	134,531
HiTZ-Aholab Basque TTS (Maider + Antton)	Phase 2: voice + prosody	CC BY 4.0	27,000

¹ Common Voice 26.0 was obtained directly from the Common Voice website. The closest HuggingFace dataset mirror is mozilla-foundation/common_voice_17_0.

Redistribution note. Because the CC BY 4.0 HiTZ-Aholab data conditions the voice of this model (Maider, Antton), any redistribution of the fine-tuned weights or generated audio must carry the HiTZ attribution and citation below.

Funding acknowledgements

The HiTZ-Aholab speech synthesis dataset was developed with funding from:

The Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia — Funded by EU — NextGenerationEU within the framework of the project ILENIA (ref. 2022/TL22/00215335), and by a grant from the Department of Culture and Language Policy of the Basque Government (IKER-GAITU project).

Citation

@dataset{itzune_zortzi_tts_onnx,
    author    = {{itzune}},
    title     = {{zortzi-tts-onnx: ONNX INT4 export of zortzi-tts (Basque TTS)}},
    year      = 2025,
    publisher = {Hugging Face},
    url       = {https://huggingface.co/itzune/zortzi-tts-onnx}
}

@inproceedings{ardila-etal-2020-common,
    title     = "Common Voice: A Massively-Multilingual Speech Corpus",
    author    = "Ardila, Rosana and Branson, Megan and Davis, Kelly and
                 Kohler, Michael and Meyer, Josh and Henretty, Michael and
                 Morais, Reuben and Saunders, Lindsay and Tyers, Francis and
                 Weber, Gregor",
    booktitle = "Proceedings of the Twelfth Language Resources and Evaluation
                 Conference (LREC)",
    month     = may, year = "2020",
    address   = "Marseille, France",
    publisher = "European Language Resources Association",
    pages     = "4218--4222",
    url       = "https://aclanthology.org/2020.lrec-1.520/"
}

@dataset{navas_hernaez_2025_17952596,
    author    = {Navas, Eva and Hernaez Rioja, Inmaculada and
                 Saratxaga, Ibon and Sanchez, Jon and
                 García Romillo, Víctor and Flores Ríos, Mariana and
                 Bellanco, Aitor},
    title     = {{HiTZ-Aholab speech synthesis dataset in Basque}},
    month     = dec, year      = 2025,
    publisher = {Zenodo},
    version   = {1.0},
    doi       = {10.5281/zenodo.17952596},
    url       = {https://doi.org/10.5281/zenodo.17952596}
}