README.md

---
language:
- en
- ru
pipeline_tag: text-to-speech
tags:
- text-to-speech
- axolotl-audio
library_name: axolotl-audio
inference: true
---

# AxolotlAudio AA-2

Multilingual neural text-to-speech from [AxolotlAudio](https://huggingface.co/AxolotlAudio).

Runtime metadata ships as a compact bundle (`assets/runtime.aa1`). Use the bundled SDK to materialize weights.

## Architecture

| Component | Description |
|-----------|-------------|
| **Backbone** | Pretrained **Qwen3-4B** (~4B params) — text planning, long-context prosody, multilingual input |
| **Acoustic head** | Hybrid autoregressive module with multi-codebook RVQ (~400M params) |
| **Codec** | Neural codec (`neural_codec.bin`) for waveform reconstruction |

The text stack reuses the **Qwen3 tokenizer** (chat template, extended audio token slots). That is intentional — we inherit Qwen3's multilingual coverage and instruction-following substrate rather than training a tokenizer from scratch.

## Design influences

AA-2 integrates ideas from several open speech / TTS research lines (not a fork of any single repo):

| Project | What we took |
|---------|--------------|
| [VITS2](https://github.com/jaywalnut310/vits) | End-to-end vocoder-less training philosophy |
| Bert-VITS2 | Multilingual BERT-style text side-conditioning |
| [GPT-SoVITS](https://github.com/RVC-Boss/GPT-SoVITS) | Two-stage text → semantic → acoustic decomposition |
| [MQTTS](https://arxiv.org/abs/2401.00438) | Multi-quantizer acoustic token modeling |
| [GPT-Fast](https://github.com/pytorch-labs/gpt-fast) | Fast secondary AR stack for residual codebooks at inference |
| [Qwen3](https://huggingface.co/Qwen) | LLM backbone + tokenizer substrate |

## Usage

```python
from axolotl_audio import load_bundle

model = load_bundle("AxolotlAudio/aa-2")
```

## Delivery markup

```
<soft> Can you hear me clearly?
<energetic> That sounds great!
```

## License

Proprietary AxolotlAudio weights.