Most speech synthesis models perform poorly with regional languages or take too long to clone a voice. Maya2 TTS, released recently, supports multilingual real-time voice cloning and uses about 1 GB of RAM. This review covers the model's key features, installation steps, and the results of local multilingual voice cloning tests based on the author's voice and English text.

General Characteristics of the Model

Maya2 TTS is an open-source model for streaming speech generation. It is designed as a conversational voice system and works with 12 languages and dialects. English is supported along with 11 regional languages.

Key parameters:

• Generation latency: 80 milliseconds. Audio begins playing almost immediately.
• Memory footprint: the quantized 4-bit version takes up 1.1 GB and runs on standard desktop RAM.
• Zero-shot voice cloning: 3–10 seconds of reference recording is enough for cloning. The model reproduces the speaker's voice in any of the 12 languages.

Architecture

The model consists of three main components:

1. Backbone based on Qwen3. It processes the input text and forms a semantic representation across the transformer layers.
2. Depth encoder. It includes 44 small transformer layers grouped into 11 stages. It produces the acoustic characteristics of the final audio.
3. Mimi decoder. It converts the result into audio with a 24 kHz sampling rate. A separate mechanism preserves the speaker's identity in the final recording.

The architecture allows switching between English and regional Indian languages within a single sentence without noticeable pauses. Cross-lingual cloning preserves the timbre of the original voice. For example, a native English speaker can obtain speech in Hindi, Marathi, or English.

Downloading and running

Two options are available:

• Official Maya2 TTS library. Downloads the full model, about 5 GB in size.
• GGUF files. Variants with 4-bit, 8-bit, and 16-bit quantization are available, along with projector files for audio encoding.

Running GGUF files requires the C++ binaries from the llama.cpp library. The steps are as follows:

1. Download all GGUF files of the chosen quantization into a local folder.
2. Download the llama.cpp binaries.
3. Run the commands in the terminal of the Google Anti-Gravity development environment.
4. Pass the main block file and the projector file as model arguments.

Testing

The test project includes three functions that are called from a common pipeline. Each function uses the Maya2 TTS launch function based on llama.cpp. The pipeline is started with a Python command from the terminal.

Task 1. Switching between English and Hindi

Task: pronounce one sentence containing an English phrase and a Hindi phrase without pauses. Text: "Hello, let me check that right away for you," followed by "order number 9842 dispatch" in Hindi.

The preset female voice Hana is used. Result: the model switched between languages without pauses, and the accent in both languages matched the norm. The phrase "Anything else I can assist you with today?" was also reproduced correctly.

Task 2. Cloning the author's voice

Task: clone the author's voice from a reference recording from the previous video and reproduce the same sentence in Hindi and Marathi. The sentence contains the words "Horizon" and "neural audio."

Results:

• Hindi. The timbre of the voice matched the reference. The last words were pronounced with an accent that is not characteristic of the author's voice.
• Marathi. The beginning was reproduced correctly, but the words at the end were jumbled, and the phrase was hard to understand. The timbre matched the reference, and the Marathi accent was assessed as accurate.

Each language was generated in a separate run, and both results were saved in separate audio files.

Task 3. Emotional sports commentary

Task: generate fast commentary in the style of a cricket commentator with Hindi inserts. The preset male voice Ritik is used. The text includes the phrases: "What an incredible finish," "Last ball before runs needed," "Bowler runs in or mid wicket," "Absolute madness here at the stadium," and "Ball goes soaring into the crowd for an unbelievable victory."

Result: the voice conveyed the emotion of the commentary and switched between Hindi and English. All three tasks, including launching the pipeline, were completed in less than 2 minutes on CPU.

Performance

Declared performance figures for different configurations:

For comparison with cloud solutions, the official data are as follows:

• Voice Arena ranking for Hindi: Maya2 TTS ranks second. Only one cloud model named in the test ranked higher (ElevenLabs or Cartesia Sonic, depending on the position in the ranking; the exact position is not specified in the source data). The competitors named are ElevenLabs and Cartesia Sonic.
• Latency: 80 milliseconds for Maya2 TTS versus an average of 120–200 milliseconds for cloud models.
• Reference recording for cloning: 3–10 seconds for Maya2 TTS versus more than 10 seconds for ElevenLabs. Other cloud solutions offer only ready-made voice profiles.
• Language support: According to the developers, Maya2 TTS supports all local languages and accents. According to the source, cloud solutions show a noticeable bias toward Western accents and a limited selection of regional dialects.

The performance and ranking figures are based on data from the developers and the source of the material. No independent verification was carried out as part of this testing.

Advantages and Limitations

Advantages:

• Zero cost. Unlimited number of generations when run locally.
• Privacy. Voice data stays in the local system and is not received by external APIs.
• Minimal latency. An 80-millisecond latency allows building voice assistants without noticeable pauses in the dialogue.

Limitations:

• Audio quality. The output format is mono 24 kHz. The format is suitable for voice assistants; studio-quality audio requires upscaling.
• Runtime environment. GGUF files require the llama.cpp library in Python. The standard Ollama environment, which is used for regular language models, is not suitable for this model.
• Punctuation. Stable output at fast speech requires correct punctuation, in particular commas in the input text.
• Quality in certain languages. In tests, Marathi and some Hindi phrases contained jumbled words and discrepancies at the end of phrases.

The model provides multilingual real-time voice synthesis on local consumer hardware with no API fees. It can be further developed for specific scenarios using the source code and roadmap described in the readme file.