Maya 2 Native

Maya 2 Native

Speaks text aloud in 11 Indian languages and English at 24 kHz, streaming audio as it is made, with 20 built-in voices and cross-lingual voice cloning. An open 2.5B text-to-speech model you can host yourself.

Made by maya-research on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Maya 2 Native is an open-weight text-to-speech model from maya-research for 11 Indian languages and English. It makes 24 kHz speech and streams the audio as it is produced. It has 20 built-in voices, supports zero-shot voice cloning and can speak one language in a voice taken from another.

Who it suits: Developers who build voice agents, phone menus, accessibility tools and narration in Indian languages and want a model they can host themselves.

Industry | Who uses it | How they use it | Example
Customer service | Voice agent builders | Speak replies in Hindi, Tamil or Telugu | A support agent that answers by voice
Telecom and banking | IVR developers | Read menus and notices aloud | A phone menu in Bengali
Accessibility | Tool makers | Read text aloud for people who prefer listening | A reader for Marathi pages
Media | Narrators and publishers | Narrate content in several languages | A Hindi voice-over for a short clip

If you need a hosted service with no setup, you can skip it, since you run the model yourself.

Where to use it:

• Python, through the inference engine's Python API
• An OpenAI-compatible speech endpoint that the engine provides
• WebSocket streaming for live voice applications

Use case | Who | What they get back
Voice agents | Developers | Streaming speech in 24 kHz audio
Narration | Creators | Speech in one of 20 built-in voices
Cross-language speech | Product teams | One voice speaking in a language other than its own

How to set it up:

• Python and pip
• About 5 GB of disk space in BF16, or about 1.3 GB at 4-bit (an estimate)
• The card's inference engine, which holds the Python API and the speech endpoint

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Set up the inference engine by following the model card, then call the Python API or the endpoint.

pip install -U huggingface_hub
hf download maya-research/maya2-native

How to use it: Send text through the Python API, the speech endpoint or WebSocket streaming and pick one of the 20 built-in voices. The engine also normalizes numbers, dates, currency and similar written forms before speaking. The card shows no run command in the part available here, so follow its quickstart.

Example: You send a short Hindi sentence that includes a date and a price. The engine reads the date and price out in words, and the model streams back 24 kHz speech in the built-in voice you chose.

For newcomers and experts: If you are new to this, listen to the card's speech samples first, then try one built-in voice. Only clone a voice that belongs to you or that you have the right to use.

The model has 2.5B parameters and is based on Qwen3-1.7B-Base. It streams mono audio in 80 ms frames. The package, Python class and configuration file keep the older names maya3, Maya3TTS and maya3_config.json. The engine adds continuous batching for serving several requests.

You send text to the engine
It normalizes numbers, dates and currency
The model makes speech in your chosen voice
Audio streams back in 80 ms frames

Good to know:

• Benchmark: on VoiceArena's Hindi real-time view the card ranks it second, and its all-model Hindi view ranks it lower
• Voice cloning: use only voices you have the right to use

Title: VoiceArena Hindi real-time Elo, from the card's table
Sonic-3.6: 1094
Maya-2-Native: 1047
Sonic-3.5: 1035
Lightning 3.1 Pro: 1024
Falcon 2: 1021
Sonic-3: 961
Bulbul-V3: 957
Grok TTS: 890

• License: Apache-2.0 for the weights and inference code
• Cost to run: no API key or paid service is named for self-hosting, only your own hardware
• Data it reaches: downloading fetches model files, and the engine reads the text you send and any voice sample you give for cloning
• Not for: cloning a voice without the right to use it

On Hugging Face, 13 November 2026

• Task: text to speech
• Parameters: 2.5B
• Weights: about 5 GB in BF16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,774
• Likes: 170
• License: apache-2.0
• Languages: as, bn, en, gu, hi, kn, ml, mr
• Based on: Qwen/Qwen3-1.7B-Base
• Quantized versions: 3
• Fine-tunes: 21
• Tags: safetensors, maya3, text-to-speech, indic, multilingual, voice-cloning, streaming, streaming-tts, multilingual-tts, indian-languages, voice-agents, speech-synthesis

Links

• Model: https://huggingface.co/maya-research/maya2-native
• Files: https://huggingface.co/maya-research/maya2-native/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download maya-research/maya2-native

Then follow the usage in its model card. Nothing in this kit runs by itself.