README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-TTS-12Hz-0.6B-Base
pipeline_tag: text-to-speech
language:
- en
tags:
- audio
- tts
- qwen3-tts
- custom-voice
- fine-tuned
---

# E3 Voice4 / Jaiden — Qwen3-TTS 0.6B direct fine-tune

This is the ear-selected **epoch-0 full fine-tune** of
[`Qwen/Qwen3-TTS-12Hz-0.6B-Base`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base)
for the E3 Voice4 single-speaker voice. The embedded speaker is exposed as
`speaker_test`; inference does not require reference audio.

The checkpoint was selected from five fixed-seed draws. The preferred output
was draw 1 (seed 43), chosen for its balance of speaker identity and expressive
range. This replaces the older Vivian/Fish-distilled weights previously on the
repository's `main` branch. Historical models remain available through the
repository's `v1`–`v5` tags and commit history.

## Quickstart

```bash
pip install -U qwen-tts soundfile
```

```python
import soundfile as sf
import torch
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "e3group-research/enderv1.0",
    device_map="cuda:0",
    dtype=torch.bfloat16,
)

torch.manual_seed(43)
wavs, sample_rate = model.generate_custom_voice(
    text="This is a test of the E3 Voice4 checkpoint.",
    speaker="speaker_test",
    language="english",
)
sf.write("voice4.wav", wavs[0], sample_rate)
```

The only configured speaker key is `speaker_test` (speaker slot 3000).

## Training recipe

- Base model: Qwen3-TTS 12 Hz 0.6B Base
- Method: single-stage full fine-tune, not LoRA or teacher distillation
- Data: 130 single-speaker speech clips, 16.56 minutes total
- Retained checkpoint: epoch 0 (65 optimizer steps)
- Learning rate: `2e-5`, with 50 warmup steps
- Objective: codec-0 cross entropy + `0.3 ×` sub-talker loss
- Embedded speaker: slot 3000, populated from the training reference voice

`history.json` records the complete eight-epoch training run. Only epoch 0 was
retained for this release; later epochs were rejected because they overfit and
could produce severe duration runaway.

## Selection measurements

For the selected seed-43 evaluation draw:

| Measurement | Value |
|---|---:|
| Speaker similarity vs. internal reference | 0.9261 |
| F0 p10–p90 range | 12.45 semitones |
| F0 median | 246.83 Hz |
| Duration | 22.78 seconds |
| Speaking rate | 187 WPM |

These are measurements for the selected generation, not aggregate benchmark
results. Generation is stochastic and results vary with text and seed.

## Integrity

- `model.safetensors` SHA-256: `54f399102bdfc0b55e5c2ffa7fd1d3eb3d68cae4a516e151cbe6df8d01e96b7a`
- `speech_tokenizer/model.safetensors` SHA-256: `836b7b357f5ea43e889936a3709af68dfe3751881acefe4ecf0dbd30ba571258`

## Limitations and responsible use

The checkpoint was selected primarily on English narration and has not been
validated equally across all languages or production workloads. It may inherit
pronunciation, hallucination, pacing, and robustness limitations from the base
model.

This is a research voice checkpoint. It is not affiliated with or endorsed by
the source speaker. Do not use it for deceptive impersonation, fraud,
harassment, or to imply endorsement. Users are responsible for obtaining any
required consent and complying with applicable privacy, publicity, copyright,
and disclosure requirements.

## Base model

- [Qwen3-TTS technical report](https://huggingface.co/papers/2601.15621)
- [Qwen3-TTS source repository](https://github.com/QwenLM/Qwen3-TTS)