What is Maya 1?

Maya 1 is an open-source speech model designed for expressive voice generation with human emotion and precise voice design. Unlike traditional text-to-speech systems that rely on pre-recorded voice libraries, Maya 1 allows you to create voices on demand using natural language descriptions.

Most voice AI systems sound clear but emotionally flat, similar to navigation systems from over a decade ago. Maya 1 changes this by generating voices that carry genuine emotion and personality. You describe the voice you want in plain English, and the model builds it instantly without requiring training data or complex parameters.

The system works by converting text into emotion-rich speech through a neural codec called SNAC. This approach generates compact audio tokens instead of raw waveforms, making it fast enough for real-time applications while maintaining high quality. The model contains 3 billion parameters and runs efficiently on a single GPU, making it accessible for both research and production use.

Maya 1 supports over 20 different emotions that you can insert directly into your text. These include laugh, cry, whisper, angry, sigh, gasp, giggle, and many more. Each emotion tag actually modifies the audio waveform, changing pitch, breath patterns, and timing to create authentic emotional expression.

Overview of Maya 1

Try Maya 1 Demo

Experience Maya 1 voice generation in action. The demo below lets you test voice design and emotion control features.

How Maya 1 Works

Maya 1 uses a decoder-only transformer architecture similar to language models like Llama, but instead of generating text tokens, it predicts audio tokens from the SNAC neural codec. This design choice makes the model efficient and fast.

Traditional speech models generate raw audio waveforms, which requires processing thousands of samples per second. Maya 1 generates compact SNAC tokens instead, with 7 tokens per audio frame. The decoder then reconstructs these tokens into 24 kHz audio. This approach dramatically reduces the sequence length, making real-time generation possible.

The SNAC codec operates at approximately 0.98 kbps, which seems extremely low but works because it uses a hierarchical structure. Multiple scales capture both fine texture and slow rhythm, creating fluid, natural-sounding speech rather than robotic output.

Training occurred in two stages. First, pretraining on internet-scale English speech data taught the model how real speech flows and how syllables connect. Second, fine-tuning on curated studio recordings with human-verified descriptions, emotion tags, accent variations, and character roles refined the model for production use.

The training data underwent strict preprocessing: 24 kHz resampling, LUFS normalization, silence trimming with voice activity detection, phrase-level alignment, and deduplication of both text and audio. Every second of data was SNAC-encoded before training, so the model learned directly from compact representations rather than waveforms.

Installation and Setup

Requirements

Before installing Maya 1, ensure you have the following:

• Python 3.8 or higher
• PyTorch installed with CUDA support
• A GPU with at least 16GB VRAM (A100, H100, or RTX 4090 recommended)
• Git LFS for downloading model weights

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 weights:

git lfs install

Step 3: Load the Model

You can load Maya 1 directly from Hugging Face using the transformers library:

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")

Step 4: Generate Voice

Create a voice description and generate speech:

# Design your voice
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
prompt = f'<description="{description}"> {text}'

# Generate 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
    )

# Decode and save audio
# (Full decoding code available in documentation)

Alternative: Clone Repository

You can also clone the model repository directly:

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

Use Cases for Maya 1

Pros and Cons

Maya 1 FAQs