---
tags:
- ml-intern
---
# Astra-TTS Architecture

Architecture design documents for Astra-TTS — a lightweight, high-quality text-to-speech system based on ZipVoice/Zipformer.

## Documents

| File | Description |
|------|-------------|
| [`model_a_slim.md`](model_a_slim.md) | **Model A** — ZipVoice naively shrunk to ~55M params. Serves as baseline. |
| [`model_b_enhanced.md`](model_b_enhanced.md) | **Model B** — ~55M params with architectural improvements (GQA, DepthSep Conv, Grouped Param Sharing, Dilated ConvNeXt, RoPE, etc.) + inference optimizations (EPSS, Midpoint ODE, SmoothCache). |
| [`benchmark_prd.md`](benchmark_prd.md) | **Benchmark PRD** — Full evaluation protocol comparing Original ZipVoice (123M) vs Model A (55M) vs Model B (55M) on LibriTTS. |

## Goal

Determine whether smart architectural changes at ~55M params can match or exceed a naive shrink, while enabling 6-8× faster inference through combined architecture + inference-time optimizations.

## Architecture Summary

| | Original ZipVoice | Model A (Slim) | Model B (Enhanced) |
|--|-------------------|---------------|-------------------|
| **Params** | 123M | ~55M | ~55M |
| **Approach** | Full size | Naive shrink | Smart redesign |
| **Key changes** | — | Smaller dims/fewer layers | GQA, DepthSep FFN, Grouped Sharing, Dilated ConvNeXt, RoPE, ConvNeXt text refinement, no NLA |
| **Inference** | Euler 16 NFE | Euler 16 NFE | Midpoint 4-step + EPSS + SmoothCache |
| **Expected speed** | 1× | ~1.5× | **~6-8×** |

## References

- ZipVoice: [arXiv:2506.13053](https://arxiv.org/abs/2506.13053)
- Zipformer: [arXiv:2310.11230](https://arxiv.org/abs/2310.11230)
- Supertonic 3: [Supertone/supertonic-3](https://huggingface.co/Supertone/supertonic-3)
- F5-TTS: [arXiv:2410.06885](https://arxiv.org/abs/2410.06885)

## License

Apache-2.0

<!-- ml-intern-provenance -->
## Generated by ML Intern

This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub.

- Try ML Intern: https://smolagents-ml-intern.hf.space
- Source code: https://github.com/huggingface/ml-intern

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Praha-Labs/Astra-TTS-Arch"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
```

For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class.