Indic Parler-TTS

For open-source Hindi and Indian language TTS, use these specific models:

1. **AI4Bharat Indic-TTS**: The premier choice for Indian languages. Built on **FastPitch** and **HiFi-GAN**, it covers Hindi and 12+ regional languages (Tamil, Telugu, Bengali, Marathi, etc.) with accurate native phonetics.
2. **Meta MMS-TTS (`facebook/mms-tts-hin`)**: Based on the **VITS** end-to-end architecture. It has dedicated checkpoints for Hindi and dozens of other Indian languages. Highly efficient and fast.
3. **Coqui XTTS-v2**: If you need zero-shot voice cloning and emotional expressiveness, XTTS-v2 supports Hindi natively out of the box.

**Serving and Real-Time Agent Stack:**
* **sherpa-onnx**: Run exportable ONNX versions of **Meta MMS-TTS** or VITS models. It is written in C++ and optimized for ultra-low latency (<150ms Time-To-First-Audio) on CPU/GPU.
* **Triton Inference Server**: Use this if deploying **AI4Bharat Indic-TTS** at scale with separated acoustic (FastPitch) and vocoder (HiFi-GAN) pipelines.

*Go with **Meta MMS via sherpa-onnx** for the fastest real-time streaming, or **AI4Bharat Indic-TTS** for the highest linguistic accuracy across Indian states.*

For Indian languages, the best open-source models are:

1. **AI4Bharat Indic-TTS (VITS / FastPitch)**: The gold standard for Indian languages, supporting 13+ languages (Hindi, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam, etc.). It outperforms global models in native accent and prosody. Checkpoints and datasets are available on Hugging Face under the `ai4bharat` organization (e.g., `ai4bharat/indic-tts-coqui-indoaryan`).
2. **Meta MMS-TTS (Massively Multilingual Speech)**: Best for sheer dialect variety, covering dozens of scheduled and non-scheduled Indian languages. Built on the VITS architecture and natively integrated into Hugging Face `transformers` (`facebook/mms-tts-[lang-code]`).
3. **Coqui XTTS-v2**: The best choice if you need real-time, zero-shot voice cloning, though its Indic support is currently limited to **Hindi**.

**Recommended Deployment Stack for an Agent:**
* **Model**: **AI4Bharat Indic-TTS** (VITS checkpoints).
* **Serving**: Convert checkpoints to **ONNX Runtime** for low-latency CPU/GPU execution, or deploy via **NVIDIA Triton Inference Server** with streaming chunks to achieve sub-300ms time-to-first-audio (TTFA).
* **Text Preprocessing**: Pair with AI4Bharat’s `indic-numeral` and `indic-transliteration` libraries to handle code-mixed (Hinglish) text and number normalization correctly.