README.md

---
license: mit
datasets:
- OpenSpeechHub/mls_eng_10k_snac_qwen
language:
- en
base_model:
- Qwen/Qwen3-0.6B
pipeline_tag: text-to-speech
library_name: transformers
---
## Overview
VyvoTTS-v0-Qwen3-0.6B is a Text-to-Speech model based on Qwen3-0.6B, trained to produce natural-sounding English speech.

- **Type:** Text-to-Speech  
- **Language:** English  
- **License:** MIT  
- **Params:** ~810M  

> **Note:** This model has a high Word Error Rate (WER) as it was trained on a 10,000-hour dataset. To improve the model's accuracy, you should use it as a pretrained base.
> I can recommend the Emilia dataset for this purpose. After the pretraining process is complete, you should perform fine-tuning for single-speaker speech.

## Usage
Below is an example of using the model with `unsloth` and `SNAC` for speech generation:

```python
from unsloth import FastLanguageModel
import torch
from snac import SNAC

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Vyvo/VyvoTTS-v0-Qwen3-0.6B",
    max_seq_length= 8192,
    dtype = None,
    load_in_4bit = False,
)
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
tokeniser_length = 151669
start_of_text = 151643
end_of_text = 151645

start_of_speech = tokeniser_length + 1
end_of_speech = tokeniser_length + 2
start_of_human = tokeniser_length + 3
end_of_human = tokeniser_length + 4
pad_token = tokeniser_length + 7

audio_tokens_start = tokeniser_length + 10
prompts = ["Hey there my name is Elise, and I'm a speech generation model that can sound like a person."]
chosen_voice = None

FastLanguageModel.for_inference(model)
snac_model.to("cpu")
prompts_ = [(f"{chosen_voice}: " + p) if chosen_voice else p for p in prompts]

all_input_ids = []
for prompt in prompts_:
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
    all_input_ids.append(input_ids)

start_token = torch.tensor([[start_of_human]], dtype=torch.int64)
end_tokens = torch.tensor([[end_of_text, end_of_human]], dtype=torch.int64)

all_modified_input_ids = []
for input_ids in all_input_ids:
    modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1)
    all_modified_input_ids.append(modified_input_ids)

all_padded_tensors, all_attention_masks = [], []
max_length = max([m.shape[1] for m in all_modified_input_ids])
for m in all_modified_input_ids:
    padding = max_length - m.shape[1]
    padded_tensor = torch.cat([torch.full((1, padding), pad_token, dtype=torch.int64), m], dim=1)
    attention_mask = torch.cat([torch.zeros((1, padding), dtype=torch.int64), torch.ones((1, m.shape[1]), dtype=torch.int64)], dim=1)
    all_padded_tensors.append(padded_tensor)
    all_attention_masks.append(attention_mask)

input_ids = torch.cat(all_padded_tensors, dim=0).to("cuda")
attention_mask = torch.cat(all_attention_masks, dim=0).to("cuda")

generated_ids = model.generate(
    input_ids=input_ids,
    attention_mask=attention_mask,
    max_new_tokens=1200,
    do_sample=True,
    temperature=0.6,
    top_p=0.95,
    repetition_penalty=1.1,
    num_return_sequences=1,
    eos_token_id=end_of_speech,
    use_cache=True
)

token_to_find = start_of_speech
token_to_remove = end_of_speech
token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)

if len(token_indices[1]) > 0:
    last_occurrence_idx = token_indices[1][-1].item()
    cropped_tensor = generated_ids[:, last_occurrence_idx+1:]
else:
    cropped_tensor = generated_ids

processed_rows = []
for row in cropped_tensor:
    masked_row = row[row != token_to_remove]
    processed_rows.append(masked_row)

code_lists = []
for row in processed_rows:
    row_length = row.size(0)
    new_length = (row_length // 7) * 7
    trimmed_row = row[:new_length]
    trimmed_row = [t - audio_tokens_start for t in trimmed_row]
    code_lists.append(trimmed_row)

def redistribute_codes(code_list):
    layer_1, layer_2, layer_3 = [], [], []
    for i in range((len(code_list)+1)//7):
        layer_1.append(code_list[7*i])
        layer_2.append(code_list[7*i+1]-4096)
        layer_3.append(code_list[7*i+2]-(2*4096))
        layer_3.append(code_list[7*i+3]-(3*4096))
        layer_2.append(code_list[7*i+4]-(4*4096))
        layer_3.append(code_list[7*i+5]-(5*4096))
        layer_3.append(code_list[7*i+6]-(6*4096))
    codes = [
        torch.tensor(layer_1).unsqueeze(0),
        torch.tensor(layer_2).unsqueeze(0),
        torch.tensor(layer_3).unsqueeze(0)
    ]
    audio_hat = snac_model.decode(codes)
    return audio_hat

my_samples = []
for code_list in code_lists:
    samples = redistribute_codes(code_list)
    my_samples.append(samples)

from IPython.display import display, Audio
if len(prompts) != len(my_samples):
    raise Exception("Number of prompts and samples do not match")
else:
    for i in range(len(my_samples)):
        print(prompts[i])
        samples = my_samples[i]
        display(Audio(samples.detach().squeeze().to("cpu").numpy(), rate=24000))

del my_samples, samples
```

## Citation

If you use this model, please cite:

```bibtex
@misc{VyvoTTS-v0-Qwen3-0.6B,
  title={VyvoTTS-v0-Qwen3-0.6B},
  author={Vyvo},
  year={2025},
  howpublished={\url{https://huggingface.co/Vyvo/VyvoTTS-v0-Qwen3-0.6B}}
}
```