Qwen3-TTS 0.6B Hindi (female voice)

Fine-tune of Qwen/Qwen3-TTS-12Hz-0.6B-Base for Hindi, one female voice (hi_f), trained on the Hindi female speaker of ai4bharat/Rasa. Single speaker, Hindi only. Trained on ai4bharat/Rasa (CC-BY-4.0). Please credit the dataset authors if you use or redistribute this model.

Training data

The Rasa Hindi female training split has 14,532 clips (26.16 h). Clips that would teach the model wrong things were removed:

• 307 clips under 1 s (too short to learn from).
• 129 clips over 20 s (too long for the 20 s generation cap).
• 492 clips where Whisper's transcript of the audio differed strongly from the text (CER over 0.5), which usually means a wrong or odd transcript, or a proper noun.

That left 13,604 clips (24.81 h). The Whisper check was only run on part of the data: 7,488 of the kept clips were not ASR-checked. A separate 300-clip test set and 150-clip validation set were held out and never used for training.

Training settings

Full fine-tune (fp32 weights), 2 epochs (two passes over the 13,604 clips), learning rate 2e-5, batch size 2 with gradient accumulation 2, seed 1234.

Usage

import torch, soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained("karan7718/qwen3-tts-0.6b-hindi", device_map="cuda:0", dtype=torch.bfloat16)
wavs, sr = model.generate_custom_voice(text="नमस्ते, आप कैसे हैं?", speaker="hi_f", language="auto", max_new_tokens=250)
sf.write("out.wav", wavs[0], sr)

Use language="auto" (Hindi is not in the base model's language list). max_new_tokens=250 is about 20 s of audio.

Results (300 held-out Rasa female test clips, Whisper large-v3 hi)

Weak spots: proper nouns (CER about 0.36 here; Whisper also struggles on them) and Alexa-style prompts (about 0.25). Not evaluated: other speakers, voice cloning, long text (over 20 s), human MOS.