Skip to content

GPA_TTS

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

GPA_TTS

README.md

GPA-TTS

Lightweight Voice-Cloning TTS for Edge Devices

Hugging Face GPA

TL;DR GPA-TTS is a standalone TTS runtime extracted from GPA, featuring zero-shot voice cloning, one of the smallest memory footprints, selectable decoder precision (INT8 / FP16 / FP32), and ready-to-run scripts.

💡 Why GPA-TTS?

Based on usage patterns from our HuggingFace Space demo, Text-to-Speech has been the most popular feature among GPA users.

While our upcoming GPA-v1.5 will be a larger and more capable unified model, we recognized the community's need for a ultra-efficient, deployment-ready TTS solution.

Thus, we distilled the TTS component from GPA into this standalone package.

GPA-TTS is designed with edge deployment in mind:

  • Compact: INT8 quantization + ONNX runtime enables one of the smallest memory footprints among open-source TTS solutions
  • Selectable Decoder Precision: Choose INT8 (smallest footprint), FP16 (balanced), or FP32 (highest quality) at runtime to match your hardware budget
  • Voice Cloning: Zero-shot cloning from a short reference audio
  • Self-Contained: No external LLM server required—everything runs locally
  • Production-Ready: REST API with voice registration and management

We hope this release serves developers building voice applications on resource-constrained devices.

Table of Contents

📥 Download

Download the model files from Hugging Face:

🤗 AutoArk-AI/GPA/GPA_TTS

The INT8 decoder is bundled by default. FP16 and FP32 decoders are provided as optional downloads for users who have extra compute headroom and want higher synthesis quality.

Package Path on HF Required
Core (INT8) GPA_TTS/GPA_TTS_INT8/ ✅
FP16 Decoder GPA_TTS/GPA_TTS_FP16/ Optional
FP32 Decoder GPA_TTS/GPA_TTS_FP32/ Optional

After downloading, your local directory should look like this:

GPA_TTS/
├── model/
│   ├── qwen_int4_ort/
│   ├── reference/
│   │   └── 038142_global_tokens.npy
│   ├── runtime_manifest.json
│   ├── spark_detokenizer_int8.onnx
│   ├── spark_detokenizer_int8.onnx.data
│   ├── spark_detokenizer_fp16.onnx          # optional
│   ├── spark_detokenizer_fp16.onnx.data     # optional
│   ├── spark_detokenizer_fp32.onnx          # optional
│   └── spark_detokenizer_fp32.onnx.data     # optional
├── voice/
│   └── spark_tokenizer_model/
│       ├── config.json
│       ├── config.yaml
│       ├── model.safetensors
│       └── wav2vec2-large-xlsr-53/
├── voices/
│   ├── items/
│   │   └── default/
│   │       ├── global_tokens.npy
│   │       └── meta.json
│   └── registry.json
├── requirements.txt
├── service.py
├── spark_tokenizer_runtime/
├── static/
├── tts_realtime_int8_light.py
├── voice_registration.py
└── voice_registry.py

⚠️ Important: All three directories (model/, voice/, voices/) must be present. If any are missing, the runtime will not start correctly.

💡 Tip: To enable FP16 or FP32 decoding, simply place the corresponding .onnx and .onnx.data files into your model/ directory alongside the INT8 artifacts. The runtime falls back to whichever precision is available.

🧹 Environment Setup

Create a Python environment and install dependencies:

python3.12 -m venv .venv_tts_onnx
source .venv_tts_onnx/bin/activate
pip install -r requirements.txt

🔊 CLI Synthesis

Run a quick smoke test with the bundled default voice:

python tts_realtime_int8_light.py \
  --text "Hello, this is a test." \
  --global-token-path voices/items/default/global_tokens.npy \
  --output-path ./hello.wav

To use a higher-precision decoder (requires the corresponding ONNX file in model/):

python tts_realtime_int8_light.py \
  --text "Hello, this is a test." \
  --global-token-path voices/items/default/global_tokens.npy \
  --decoder-precision fp16 \
  --output-path ./hello_fp16.wav

Expected output:

  • hello.wav generated in the current directory
  • Memory checkpoints and synthesis summary printed to console

💡 Tip: If you encounter missing-file errors, verify that the HF assets were restored to the exact paths shown in the directory structure above.

🌐 Local Service

Start the FastAPI-based local server:

python -m uvicorn service:app --host 127.0.0.1 --port 8021

Once running, access:

Page URL
Home http://127.0.0.1:8021/
Chinese Demo http://127.0.0.1:8021/static/demo_page_cn.html
English Demo http://127.0.0.1:8021/static/demo_page_en.html

📡 API Reference

Health Check

curl http://127.0.0.1:8021/api/health

Memory Usage

Returns current and peak memory usage of the service process—useful for monitoring resource consumption on edge devices.

curl http://127.0.0.1:8021/api/memory

List Registered Voices

curl http://127.0.0.1:8021/api/voices

Text-to-Speech

curl -X POST http://127.0.0.1:8021/api/tts \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, this is a test.",
    "voice_name": "default",
    "decoder_precision": "int8",
    "temperature": 0.3,
    "repetition_penalty": 1.2,
    "max_new_tokens": 512,
    "do_sample": false
  }'

💡 Tip: Set decoder_precision to "fp16" or "fp32" for higher synthesis quality at the cost of more memory. Omit the field or use "int8" for the smallest footprint.

Register Voice (Local Path)

curl -X POST http://127.0.0.1:8021/api/voices/register-path \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-voice",
    "audio_path": "./reference_audio/sample.wav",
    "overwrite": false
  }'

Register Voice (File Upload)

curl -X POST http://127.0.0.1:8021/api/voices/register-upload \
  -F "name=my-voice" \
  -F "overwrite=false" \
  -F "audio=@./reference_audio/sample.wav"

🎛️ Decoder Precision

GPA-TTS ships with three decoder variants. The runtime automatically falls back to whichever precision is available if the requested one is not found.

Variant File Memory Quality Recommended For
INT8 spark_detokenizer_int8.onnx Lowest Good Edge / mobile / memory-constrained
FP16 spark_detokenizer_fp16.onnx Medium Better Desktop / server with moderate resources
FP32 spark_detokenizer_fp32.onnx Highest Best Server with ample resources; quality-critical

To switch precision:

  • CLI: --decoder-precision fp16
  • API: "decoder_precision": "fp16" in the JSON body
  • Web UI: Select from the "Decoder Precision" dropdown

🙏 Acknowledgements

Special thanks to @gh503 for reporting the missing GPA-TTS runtime dependency and sharing the fix in Issue #19.

🔗 Related