dhee-indic-f5

Dhee-Indic-F5

A zero-shot, voice-cloning text-to-speech model for Indian languages and Indian-accented English, fine-tuned from ai4bharat/IndicF5. Give it a short reference clip (audio + its transcript) and target text, and it speaks the target in the reference speaker's voice.

What this adds over the base IndicF5

IndicF5 is a strong flow-matching TTS backbone for 11 Indian languages. This fine-tune keeps that foundation and focuses on:

• Indian-accented English as a first-class target, alongside the Indic languages — useful for code-mixed and English-heavy Indian content that the base model handles less naturally.
• Steadier zero-shot cloning from short references — more consistent speaker identity and prosody when the reference clip is only a few seconds long.
• Cleaner handling of code-mixed input (English words inside an Indic sentence), which is common in real Indian text.

Everything else — the architecture, the inference API, and the reference-clip conditioning — is unchanged from IndicF5, so existing IndicF5 code works as-is by swapping the model id.

Supported languages

Assamese, Bengali, English (Indian), Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu.

Usage

This model uses the IndicF5 interface. Provide a reference audio clip, the exact transcript of that clip, and the text you want spoken.

from transformers import AutoModel
import numpy as np, soundfile as sf

repo = "dheeyantra/dhee-indic-f5"
model = AutoModel.from_pretrained(repo, trust_remote_code=True)

audio = model(
    "नमस्ते, आप कैसे हैं?",                 # text to synthesize
    ref_audio_path="ref.wav",              # a few seconds of the target speaker
    ref_text="This is the reference transcript.",  # exact transcript of ref.wav
)

if audio.dtype == np.int16:
    audio = audio.astype(np.float32) / 32768.0
sf.write("out.wav", np.array(audio, dtype=np.float32), samplerate=24000)

Tips

• Keep the reference clip clean and 3–10 seconds long; ref_text must match it exactly.
• For best results, match the reference speaker's language/accent to your target text (e.g. an Indian-English reference for Indian-English output).

Limitations

• Zero-shot cloning quality tracks the reference clip: noisy or very short references degrade output.
• Not intended for high-fidelity singing or heavy non-speech vocalizations.
• Do not use it to imitate a real person's voice without their consent.

Licence

Released under CC-BY-NC-4.0 — free to use, share, and adapt for non-commercial purposes with attribution.

This model inherits non-commercial terms from its lineage: it is fine-tuned from ai4bharat/IndicF5, which builds on F5-TTS (trained on the Emilia dataset, CC-BY-NC-4.0) and IIT-Madras Indic TTS data (non-commercial). The non-commercial restriction therefore carries through to this model.

For commercial licensing, contact [email protected].

Attribution

• Base model: ai4bharat/IndicF5 (MIT)
• Vocoder: Vocos (MIT)

Citation

@misc{dhee_indic_f5,
  title  = {Dhee-Indic-F5: Indian-language voice-cloning TTS},
  author = {DheeYantra Research Labs},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/dheeyantra/dhee-indic-f5}}
}