Qwen3-1.7B-Multilingual-TTS

Continue pretraining Qwen/Qwen3-1.7B-Base on Multilingual Voice Conversion and TTS.

1. Use neucodec as speech detokenizer, 50 TPS, output in 24k sample rate.
2. Multi-speaker multilingual Voice Conversion up to 25.5B tokens.
3. Multi-speaker multilingual TTS up to 5B tokens.
4. Flash Attention 3 10k context length multipacking.
5. Liger Kernel for swiglu, rms_norm and fused_linear_cross_entropy.

WanDB at https://wandb.ai/huseinzol05/Qwen-Qwen3-1.7B-Base-multilingual-tts-neucodec

We released better version at Scicom-intl/Multilingual-TTS-1.7B-Base.

How to

import soundfile as sf
import torch
import torchaudio
from transformers import AutoTokenizer, AutoModelForCausalLM
from neucodec import NeuCodec
import re

model_name = "malaysia-ai/Qwen3-1.7B-Multilingual-TTS"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto").to('cuda')
codec = NeuCodec.from_pretrained("neuphonic/neucodec")
_ = codec.eval().to('cuda')

TTS

text = "Hello! how come I help you? 你好！有什么可以帮你的吗？வணக்கம்! நான் உங்களுக்கு எப்படி உதவுவது? Bonjour! Comment puis-je vous aider ? Xin chào! Tôi có thể giúp gì cho bạn? こんにちは！どうしてお手伝いしましょうか？안녕하세요! 어떻게 도와드릴까요?"
prompt = f"<|im_start|>jenny_tts_dataset_audio_jenny: {text}<|speech_start|>"

inputs = tokenizer(prompt,return_tensors="pt", add_special_tokens=True).to('cuda')

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=2048,
        do_sample=True,
        temperature=0.6,
        repetition_penalty=1.15,
    )

generated_text = tokenizer.decode(outputs[0], skip_special_tokens=False)
audio_tokens = re.findall(r'<\|s_(\d+)\|>', generated_text.split('<|speech_start|>')[1])
audio_tokens = [int(token) for token in audio_tokens]
audio_codes = torch.tensor(audio_tokens)[None, None]

with torch.no_grad():
    audio_waveform = codec.decode_code(audio_codes.cuda())

sf.write('7-languages.mp3', audio_waveform[0, 0].cpu(), 24000)

You can check the audio 7-languages.mp3.

• You can pick any speaker name from malaysia-ai/Multilingual-TTS.
• Not bad from 0.35 epoch model.

Voice Conversion

import librosa

y, sr = librosa.load('jenny.wav', sr = 16000)
with torch.no_grad():
    codes = codec.encode_code(torch.tensor(y)[None, None])
tokens = ''.join([f'<|s_{i}|>' for i in codes[0, 0]])
prompt = f"<|im_start|>I wonder if I shall ever be happy enough to have real lace on my clothes and bows on my caps.<|speech_start|>{tokens}<|im_end|><|im_start|>Hello, how come I help you, 你好, 有什么可以帮你的吗, வணக்கம், நான் உங்களுக்கு எப்படி உதவுவது, bonjour, comment puis-je vous aider.<|speech_start|>"

inputs = tokenizer(prompt,return_tensors="pt", add_special_tokens=True).to('cuda')

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=2048,
        do_sample=True,
        temperature=0.6,
        repetition_penalty=1.15,
    )

generated_text = tokenizer.decode(outputs[0], skip_special_tokens=False)
audio_tokens = re.findall(r'<\|s_(\d+)\|>', generated_text.split('<|speech_start|>')[-1])
audio_tokens = [int(token) for token in audio_tokens]
audio_codes = torch.tensor(audio_tokens)[None, None]

with torch.no_grad():
    audio_waveform = codec.decode_code(audio_codes.cuda())

sf.write('jenny-4-languages.mp3', audio_waveform[0, 0].cpu(), 24000)

You can check the audio jenny-4-languages.mp3.

• Not too great, we need to trim the silents first before convert to audio tokens, the model tends to generate long silents.

Source code

Source code at https://github.com/malaysia-ai/cooking/tree/main/qwen-tts