Extremely poor and unintelligible Hindi output despite correct reference audio and transcript
I tested IndicF5 specifically for Hindi TTS and zero-shot voice cloning, but unfortunately the generated speech quality was extremely poor.
I used:
Hindi target text in Devanagari
Clean reference WAV audio
Exact transcript matching the reference audio
24 kHz mono reference audio
NVIDIA RTX 5070 Laptop GPU
CUDA-enabled PyTorch
The documented AutoModel.from_pretrained("ai4bharat/IndicF5", trust_remote_code=True) inference method
After resolving the installation and dependency issues, the model loaded and inference successfully ran. However, the actual generated Hindi speech was largely unintelligible.
This was not simply a voice-similarity problem. In my tests, even basic Hindi words were frequently pronounced incorrectly or generated as speech that did not clearly correspond to the supplied text. The output was not usable for narration.
My reference transcript was ordinary Hindi, for example:
“हमारा प्यारा और अनोखा देश भारत और उससे भी अनोखे हमारे दुश्मन…”
I expected the generated speech at minimum to reproduce the supplied Hindi text intelligibly, even if speaker similarity was imperfect.
I would appreciate clarification on:
Is this a known problem with Hindi inference?
Is there a specific recommended reference-audio duration?
Are there particular reference-audio characteristics required for reliable Hindi generation?
Is there a recommended Transformers/F5-TTS/PyTorch configuration that produces the intended results?
Are there official Hindi samples and reference prompts that can be used to reproduce the advertised quality?
I am happy to provide a minimal reproducible example if useful.
At the moment, based on my testing, the Hindi output is unfortunately not reliable enough for production use.
you can try the for IIT madras hindhi TTS model which is good for realt time inference