Thiva-indic by Cortiqa (FP16 Studio Quality)
Thiva-indic is a high-fidelity, on-device Text-to-Speech (TTS) model optimized for the Cortiqa Voice Studio ecosystem. It provides clear, natural, and expressive synthetic speech across 17 Indian languages and English.
The model is optimized in FP16 half-precision, reducing memory footprint by over 50 percent while preserving original studio acoustics.
Key Features
- 17 Languages in One Model: Generates speech in 16 major Indian languages plus English without separate voice packs.
- Studio-Grade Audio: 44.1 kHz clean sampling rate with natural breathing and human-like intonation.
- On-Device and Offline Ready: Engineered for local edge/mobile devices (Android, iOS, laptops) with zero server dependencies.
- Free for Commercial Use: Released under the permissive Apache 2.0 License.
Supported Languages
| # | Language | Native Script | Code |
|---|---|---|---|
| 1 | Hindi | हिन्दी | hi |
| 2 | Bengali | বাংলা | bn |
| 3 | Tamil | தமிழ் | ta |
| 4 | Telugu | తెలుగు | te |
| 5 | Marathi | मराठी | mr |
| 6 | Gujarati | ગુજરાતી | gu |
| 7 | Kannada | ಕನ್ನಡ | kn |
| 8 | Malayalam | മലയാളം | ml |
| 9 | Odia | ଓଡ଼ିଆ | or |
| 10 | Punjabi | ਪੰਜਾਬੀ | pa |
| 11 | Assamese | অসমীয়া | as |
| 12 | Urdu | اردو | ur |
| 13 | Kashmiri | कॉशुर / کٲشُر | ks |
| 14 | Nepali | नेपाली | ne |
| 15 | Sanskrit | संस्कृतम् | sa |
| 16 | Sindhi | سنڌي | sd |
| 17 | English | English | en |
Quickstart and Inference
import torch
import soundfile as sf
from parler_tts import ParlerTTSForConditionalGeneration
from transformers import AutoTokenizer
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
model = ParlerTTSForConditionalGeneration.from_pretrained('cortiqa/Thiva-indic').to(device).half()
tokenizer = AutoTokenizer.from_pretrained('cortiqa/Thiva-indic')
description_tokenizer = AutoTokenizer.from_pretrained(model.config.text_encoder._name_or_path)
prompt = 'नमस्ते! कोर्टिका के थिवा-इंडिक मॉडल में आपका स्वागत है।'
description = 'Divya speaks with a clear, pleasant, natural and articulate voice at a steady pace in a quiet studio.'
desc_inputs = description_tokenizer(description, return_tensors='pt').to(device)
prompt_inputs = tokenizer(prompt, return_tensors='pt').to(device)
with torch.no_grad():
generation = model.generate(
input_ids=desc_inputs.input_ids,
attention_mask=desc_inputs.attention_mask,
prompt_input_ids=prompt_inputs.input_ids,
prompt_attention_mask=prompt_inputs.attention_mask,
temperature=0.8,
do_sample=True,
)
audio_arr = generation.cpu().float().numpy().squeeze()
sf.write('output.wav', audio_arr, model.config.sampling_rate)
print('Audio generated: output.wav')
Voice Directing and Style Prompts
Control speaker style, gender, pacing, and mood using the description string:
- Clean Studio Female: 'Divya speaks with a clear, warm, pleasant and articulate voice at a steady pace in a quiet studio with very clean audio.'
- Punchy Studio Male: 'Rohit speaks with an energetic, modern, confident and clear male voice in a professional recording booth.'
- Formal News Narration: 'A female speaker delivers an authoritative, clear and precise news narration with deliberate pacing and zero background noise.'
Attribution and Credits
Thiva-indic is built upon and optimized from ai4bharat/indic-parler-tts developed by AI4Bharat (IIT Madras).
- Original Project: AI4Bharat Indic-Parler-TTS
- Institution: Indian Institute of Technology Madras (IIT Madras)
- License: Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
License
This model is licensed under the Apache License, Version 2.0. Free to use, modify, distribute, and integrate into commercial applications.