What is Maya 1?

Maya 1 is a speech model designed for expressive voice generation with human emotion and precise voice design. This open source text-to-speech system enables you to create natural voices that convey feelings, personalities, and specific characteristics through simple text descriptions.

Unlike traditional text-to-speech systems that produce flat, robotic voices, Maya 1 generates speech with emotional depth. You can describe a voice in natural language, add emotions at specific points in your text, and produce audio that sounds authentic and engaging. The model supports over 20 different emotions including laughter, crying, whispering, anger, sighing, gasping, and many more.

Maya 1 operates as a 3 billion parameter model built on a Llama-style transformer architecture. It generates audio using the SNAC neural codec, which produces high-quality 24 kHz audio at approximately 0.98 kbps. This makes it suitable for real-time streaming applications where low latency and bandwidth efficiency matter.

The model is released under the Apache 2.0 license, meaning you can use it freely for commercial and personal projects. You can deploy it on your own infrastructure, customize it for specific use cases, and avoid per-second fees that come with proprietary voice AI services.

Overview of Maya 1

Try Maya 1 Demo

Experience Maya 1 in action. The interactive demo below lets you test voice generation with different descriptions and emotions. Try creating various character voices and see how the model responds to your inputs.

Installation and Setup

Requirements

Before installing Maya 1, ensure you have the following:

• Python 3.8 or higher
• CUDA-capable GPU with 16GB+ VRAM (A100, H100, or RTX 4090 recommended)
• PyTorch installed with CUDA support
• Git LFS for downloading model files

Step 1: Install Dependencies

Install the required Python packages:

pip install torch transformers snac soundfile

Step 2: Install Git LFS

Git LFS is needed to download the model files:

git lfs install

Step 3: Download the Model

Clone the model repository from Hugging Face:

git clone https://huggingface.co/maya-research/maya1

Alternatively, you can load the model directly in Python without cloning:

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("maya-research/maya1")

Step 4: Basic Usage Example

Here's a simple example to generate voice with emotions:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from snac import SNAC
import soundfile as sf

# Load the model
model = AutoModelForCausalLM.from_pretrained(
    "maya-research/maya1", 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("maya-research/maya1")

# Load SNAC audio decoder
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().to("cuda")

# Design your voice with natural language
description = "Realistic male voice in the 30s age with american accent. Normal pitch, warm timbre, conversational pacing."
text = "Hello! This is Maya 1 <laugh> the best open source voice AI model with emotions."

# Create prompt with voice design
prompt = f'<description="{description}"> {text}'

# Generate emotional speech
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
with torch.inference_mode():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=500, 
        temperature=0.4, 
        top_p=0.9, 
        do_sample=True
    )

# Extract and decode SNAC tokens to audio
generated_ids = outputs[0, inputs['input_ids'].shape[1]:]
snac_tokens = [t.item() for t in generated_ids if 128266 <= t <= 156937]

# Process tokens and generate audio
frames = len(snac_tokens) // 7
codes = [[], [], []]
for i in range(frames):
    s = snac_tokens[i*7:(i+1)*7]
    codes[0].append((s[0]-128266) % 4096)
    codes[1].extend([(s[1]-128266) % 4096, (s[4]-128266) % 4096])
    codes[2].extend([(s[2]-128266) % 4096, (s[3]-128266) % 4096, (s[5]-128266) % 4096, (s[6]-128266) % 4096])

codes_tensor = [torch.tensor(c, dtype=torch.long, device="cuda").unsqueeze(0) for c in codes]
with torch.inference_mode():
    audio = snac_model.decoder(snac_model.quantizer.from_codes(codes_tensor))[0, 0].cpu().numpy()

# Save audio output
sf.write("output.wav", audio, 24000)
print("Voice generated successfully! Play output.wav")

Step 5: Production Deployment with vLLM

For production deployments with real-time streaming, use vLLM integration. This enables automatic prefix caching for repeated voice descriptions, WebAudio ring buffer integration, multi-GPU scaling support, and sub-100ms latency for real-time applications. Download the vLLM streaming inference script from the model repository for production setups.

Use Cases for Maya 1

How Maya 1 Works

Maya 1 uses a 3 billion parameter decoder-only transformer architecture, similar to Llama models, but trained to predict SNAC neural codec tokens instead of text tokens. The generation process follows a clear pipeline: you provide a voice description and text, the model tokenizes the input, generates SNAC audio code tokens (7 tokens per frame), and then decodes these tokens into 24 kHz audio.

The SNAC neural codec uses a multi-scale hierarchical structure operating at approximately 12, 23, and 47 Hz. This keeps autoregressive sequences compact, enabling real-time streaming at about 0.98 kbps while maintaining high audio quality. The codec's efficiency is what makes real-time applications possible.

Training involved two main phases. First, pretraining on an internet-scale English speech corpus provided broad acoustic coverage and natural coarticulation patterns. Second, supervised fine-tuning on a curated dataset of studio recordings included human-verified voice descriptions, 20+ emotion tags per sample, multi-accent English coverage, and character and role variations.

The data pipeline ensured quality through 24 kHz mono resampling with -23 LUFS normalization, voice activity detection with silence trimming, forced alignment for clean phrase boundaries, and both text and audio deduplication. This rigorous preprocessing contributed to the model's ability to generate consistent, high-quality voices.

The voice design system uses XML-attribute format for descriptions, which proved superior to other approaches during development. Formats like colon-separated descriptions or key-value tags either caused format drift, were too rigid, or created token bloat. The XML-attribute format allows natural language descriptions that are robust and scalable.

Pros and Cons

Maya 1 FAQs