Tharshan/indicf5_hindi-english_code_switch

IndicF5-Hinglish (repackaged)

Hindi-English code-switched TTS. This is a repackaging of itself a fine-tune of ai4bharat/IndicF5.

The weights are unchanged. What's different: vocabulary, model code and a default reference voice ship with this repo, so loading needs no gated downloads and no f5_tts install.

Usage

from transformers import AutoModel
import soundfile as sf

model = AutoModel.from_pretrained(
    "Tharshan/indicf5_hindi-english_code_switch", trust_remote_code=True)

audio, sr = model.generate("मैं आज office जा रहा हूँ, morning में एक important meeting है।")
sf.write("out.wav", audio, sr)

Your own voice (F5 is a zero-shot cloner; the transcript must match the clip):

audio, sr = model.generate(text, ref_audio="my_voice.wav", ref_text="exact transcript")

pip install -r requirements.txt

Reference voice matters more than you'd expect

Same weights, same sentences, only the prompt clip changed:

Reference	Hindi CER (base)	Hindi CER (this model)
Punjabi prompt	0.147	0.347
Hinglish prompt (bundled)	0.021	0.011

Use a reference whose language and accent resemble your target text.

Evaluation

Whisper large-v3, bundled Hinglish reference, CER against input text:

Text type	base IndicF5	this model
Pure Hindi	0.021	0.011
Pure English	0.780	0.220

English improves substantially; Hindi is not degraded.

Code-switched text: raw CER is unreliable and ranks the two models backwards. Whisper writes correctly-pronounced English words in Devanagari (morning → मॉर्णिंग), scored as errors against Latin-script references. By transcript inspection this model renders the English words correctly while base produces gibberish, despite base scoring "better". Use a transliteration-aware scorer.

Details

Base	ai4bharat/IndicF5 (F5-TTS DiT, 0.3B)
Fine-tuning data	OpenSLR-104 Hindi-English
Steps	32,990
Vocab	2546 chars, unchanged from IndicF5
Sample rate	24 kHz

Training audio was ASR-corpus material resampled to 24 kHz, so high-frequency detail is limited by the sources. Expect a slightly duller timbre than base.

Credits

• Base model and vocab.txt: ai4bharat/IndicF5 — MIT, subject to AI4Bharat's terms of use
• Model code vendored from F5-TTS — MIT
• Bundled reference voice generated with Sarvam AI bulbul:v3

Do not clone voices without the speaker's consent.