What is Maya1 TTS Model?

Maya1 is an open-source speech model designed for expressive voice generation with rich human emotion and precise voice design. Built by Maya Research, this model represents a significant step forward in making high-quality voice AI accessible to everyone.

This model allows you to create natural-sounding voices by simply describing them in plain language. Instead of working with complex parameters or technical settings, you describe the voice as if you were briefing a voice actor. The model handles over 20 different emotions including laughter, crying, whispering, anger, sighing, and gasping, bringing a human touch to synthetic speech.

Maya1 runs on a single GPU and uses a 3-billion parameter architecture based on the Llama transformer model. It generates audio at 24 kHz quality and supports real-time streaming, making it suitable for production environments where low latency matters.

Overview of Maya1

Why Maya1 is Different

Most voice AI tools today are either closed-source services that charge per second of audio generated, or open-source models that lack emotional range and natural voice control. Maya1 bridges this gap by offering production-quality emotional speech synthesis with full transparency and no usage fees.

The model stands out because it accepts natural language descriptions for voice design. You can specify age, accent, pitch, timbre, pacing, and character traits in plain English. This approach makes voice creation intuitive for developers, content creators, and researchers who need expressive speech without technical audio engineering knowledge.

Maya1 supports inline emotion tags that let you place emotional expressions exactly where they belong in your text. For example, you can insert laughter mid-sentence or add a whisper for dramatic effect. This granular control over emotional delivery creates more natural and engaging speech output.

The model uses the SNAC neural codec for audio generation, which enables real-time streaming at approximately 0.98 kilobits per second. This efficient encoding makes Maya1 practical for applications like voice assistants, interactive agents, game characters, and live content generation where latency matters.

Technical Architecture

Maya1 uses a 3-billion parameter decoder-only transformer based on the Llama architecture. Rather than predicting raw audio waveforms, the model generates SNAC neural codec tokens. These tokens represent audio in a compressed hierarchical format that enables efficient streaming and generation.

The SNAC codec uses a multi-scale hierarchical structure operating at approximately 12, 23, and 47 Hz. This structure keeps autoregressive sequences compact while maintaining audio quality. Each audio frame requires 7 tokens, making the generation process efficient enough for real-time applications.

The model was pretrained on an internet-scale English speech corpus to learn broad acoustic patterns and natural coarticulation. After pretraining, supervised fine-tuning used a curated dataset of studio recordings with human-verified voice descriptions, over 20 emotion tags per sample, multi-accent English coverage, and character variations.

The training pipeline includes 24 kHz mono resampling with loudness normalization, voice activity detection for silence trimming, forced alignment for clean phrase boundaries, text deduplication using MinHash-LSH, audio deduplication with Chromaprint, and SNAC encoding with 7-token frame packing.

Installation and Setup

Getting started with Maya1 requires installing a few Python packages and loading the model from the Hugging Face model hub. The process takes just a few minutes on a system with the appropriate GPU.

Requirements

Install the necessary Python packages:

Loading the Model

You can load Maya1 directly from Hugging Face or clone the repository:

# Load directly in Python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

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

Quick Start Example

Generate your first emotional speech with this example:

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

# Load models
model = AutoModelForCausalLM.from_pretrained(
    "maya-research/maya1",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("maya-research/maya1")
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().to("cuda")

# Design your voice
description = "Realistic male voice in the 30s age with american accent. Normal pitch, warm timbre, conversational pacing."
text = "Hello! This is Maya1 <laugh> the best open source voice AI model with emotions."

# Generate speech
prompt = f'<description="{description}"> {text}'
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
    )

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

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

Use Cases for Maya1

How Maya1 Compares

Frequently Asked Questions