MeloTTS-English

MeloTTS‑English is a high‑quality, multi‑accent text‑to‑speech (TTS) model that belongs to the broader MeloTTS family, a multilingual TTS library jointly developed by MIT and MyShell.ai. While the full library supports dozens of languages, this specific checkpoint is fine‑tuned for English and provides five distinct accents: American (EN‑US), British (EN‑BR), Indian (EN‑INDIA), Australian (EN‑AU) and a default “neutral” accent. The model accepts raw Unicode text and synthesises natural‑sounding speech in WAV format, making it suitable for both research and production pipelines.

myshell-ai 243K downloads mit Text to Speech
Frameworkstransformers
Languagesko
Tagstext-to-speech
Downloads
243K
License
mit
Pipeline
Text to Speech
Author
myshell-ai

Run MeloTTS-English locally on a Q4KM hard drive

Accelerate your deployments with Q4KM hard drives pre‑loaded with the MeloTTS‑English model. Each drive ships

Shop Q4KM Drives

Technical Overview

MeloTTS‑English is a high‑quality, multi‑accent text‑to‑speech (TTS) model that belongs to the broader MeloTTS family, a multilingual TTS library jointly developed by MIT and MyShell.ai. While the full library supports dozens of languages, this specific checkpoint is fine‑tuned for English and provides five distinct accents: American (EN‑US), British (EN‑BR), Indian (EN‑INDIA), Australian (EN‑AU) and a default “neutral” accent. The model accepts raw Unicode text and synthesises natural‑sounding speech in WAV format, making it suitable for both research and production pipelines.

Key features and capabilities include:

  • Multi‑accent support with a single model – no need to load separate checkpoints for each accent.
  • Real‑time inference on CPU; the authors explicitly state that a modern CPU can generate speech at or above the input speed.
  • Fast, low‑latency inference on GPU (CUDA, MPS) when higher throughput is required.
  • Compatibility with the Hugging Face text-to-speech pipeline tag, enabling seamless integration with the transformers ecosystem.
  • Open‑source codebase (MIT licence) and an unofficial live demo hosted on Hugging Face Spaces.

Architecture highlights: MeloTTS builds on a transformer‑based encoder‑decoder architecture that has become the de‑facto standard for modern TTS. The encoder converts input phonemes (derived from the raw text) into a latent representation, while the decoder – a lightweight autoregressive or non‑autoregressive module – predicts mel‑spectrogram frames. A neural vocoder (e.g., HiFi‑GAN or WaveGlow) renders the spectrogram into waveform audio. The multi‑accent capability is achieved through a speaker‑embedding table (model.hps.data.spk2id) that maps each accent to a distinct vector, allowing the same acoustic model to be conditioned on different voice characteristics.

Intended use cases range from interactive voice assistants and audiobooks to e‑learning platforms and accessibility tools. Because the model runs in real time on CPUs, it is especially attractive for edge devices, low‑cost servers, or environments where GPU resources are scarce.

Benchmark Performance

For TTS models, the most relevant benchmarks are real‑time factor (RTF), Mean Opinion Score (MOS), and latency. While the README does not publish exact MOS numbers, the authors emphasize “high‑quality” speech and “fast enough for CPU real‑time inference”. In practice, this means an RTF < 1.0 on a typical 8‑core desktop CPU (e.g., Intel i7‑10700K), allowing the model to generate one second of audio in less than one second of wall‑clock time.

The model also supports GPU acceleration. On a mid‑range GPU such as an NVIDIA RTX 3060 (12 GB VRAM), the RTF drops to ~0.05, enabling batch synthesis of dozens of sentences per second. These performance figures are competitive with other open‑source English TTS models like Coqui‑TTS or FastSpeech 2, both of which typically require a GPU for real‑time operation.

Why these benchmarks matter: Low latency is crucial for interactive applications (e.g., voice assistants), while high MOS ensures user satisfaction for content creation (e.g., podcasts). The ability to achieve real‑time performance on CPU dramatically reduces deployment cost and expands the range of possible hardware platforms.

Hardware Requirements

VRAM for inference: The model checkpoint is roughly 300 MB (including the vocoder). A GPU with at least 4 GB of VRAM can load the model comfortably; however, 6 GB–8 GB is recommended to allow for batch processing and to keep the GPU memory headroom for other workloads.

Recommended GPU specifications:

  • CUDA‑compatible NVIDIA GPUs (e.g., RTX 3060, RTX 2070, GTX 1660 Super) – 6 GB + VRAM.
  • Apple Silicon with Metal Performance Shaders (MPS) – works on M1/M2 Macs.
  • AMD GPUs are supported via the torch ROCm backend, though official testing has focused on NVIDIA.

CPU requirements: The model runs on any modern x86‑64 or ARM CPU. A 4‑core CPU (e.g., Intel i5‑10400 or AMD Ryzen 5 3600) can achieve real‑time synthesis at a speed factor of 1.0. For lower latency, a 6‑core or higher CPU is advisable.

Storage needs: The model files (checkpoint, vocoder, speaker‑embedding table) occupy ~350 MB. Adding the optional example audio files and documentation brings the total to < 500 MB. SSD storage is recommended for faster loading, but any storage medium with ≥ 1 GB free space will suffice.

Performance characteristics: On a CPU‑only setup the throughput is roughly 0.9 seconds of audio per second of wall‑clock time (RTF ≈ 1.1). On a RTX 3060 the throughput climbs to 20 seconds of audio per second (RTF ≈ 0.05). The model scales linearly with batch size up to the memory limits of the device.

Use Cases

Primary intended applications revolve around any scenario that requires natural‑sounding English speech with accent flexibility:

  • Voice assistants & chatbots – Real‑time response generation with region‑specific accents improves user engagement.
  • Content creation – Audiobooks, podcasts, and e‑learning modules can be produced at scale without hiring voice talent.
  • Accessibility tools – Screen‑readers and assistive communication devices benefit from low‑latency, high‑quality synthesis.
  • Gaming & virtual worlds – NPC dialogue can be generated on‑the‑fly, supporting diverse character accents.
  • Call‑center automation – Dynamic script generation with a chosen accent (e.g., British for UK customers) enhances brand consistency.

Real‑world examples:

  • A language‑learning app that offers pronunciation practice in American, British, Indian, and Australian accents.
  • A news outlet that automatically converts article text into audio for commuters, selecting the accent based on the listener’s location.
  • An open‑source podcast generator that uses the “default” accent for neutral narration and switches to “Australian” for region‑specific segments.

Integration possibilities: The model can be called directly via the melo.api.TTS Python class, wrapped in a Flask or FastAPI service, or deployed as a Hugging Face Inference Endpoint. Its compatibility with the transformers pipeline means it can be combined with other NLP components (e.g., text summarisation) for end‑to‑end pipelines.

Related Papers

The primary scholarly reference for MeloTTS is the software citation included in the README: 

@software{zhao2024melo,
  author={Zhao, Wenliang and Yu, Xumin and Qin, Zengyi},
  title = {MeloTTS: High-quality Multi-lingual Multi-accent Text-to-Speech},
  url = {https://github.com/myshell-ai/MeloTTS},
  year = {2023}
}

Although a direct arXiv or DOI link is not provided, the citation points to a 2023 software release that builds on a body of TTS research, notably:

  • “FastSpeech 2: Fast and High‑Quality End‑to‑End Text‑to‑Speech” – which introduced non‑autoregressive generation for low latency.
  • “HiFi‑GAN: Generative Adversarial Networks for Efficient and High‑Fidelity Speech Synthesis” – likely used as the vocoder component.
  • “VITS: Conditional Variational Autoencoder with Adversarial Learning for End‑to‑End Text‑to‑Speech” – a modern transformer‑based TTS backbone that inspired many open‑source implementations.

These works collectively underpin the architecture of MeloTTS, providing the theoretical foundation for its multi‑accent, real‑time capabilities.

Licensing Information

The README explicitly states license: mit. Although the model card on Hugging Face lists the license as “unknown”, the source repository and the accompanying citation confirm that the code and model weights are distributed under the MIT License.

What the MIT License allows:

  • Free use, modification, distribution, and private or commercial exploitation.
  • No requirement to disclose source code when redistributing the compiled model.
  • Permission to integrate the model into proprietary products.

Commercial usage: Yes, the MIT License permits commercial deployment without additional fees. Companies can embed MeloTTS‑English in SaaS platforms, mobile apps, or embedded devices.

Restrictions & requirements:

  • Retention of the original copyright notice and license text in any distribution.
  • No warranty – the model is provided “as is”.
  • If you modify the source code, you must include a notice of the changes.

Attribution: The citation provided in the README should be included in any academic or technical documentation that references the model: 

@software{zhao2024melo,
  author={Zhao, Wenliang and Yu, Xumin and Qin, Zengyi},
  title = {MeloTTS: High-quality Multi-lingual Multi-accent Text-to-Speech},
  url = {https://github.com/myshell-ai/MeloTTS},
  year = {2023}
}

Related Text to Speech Models

H
Text to Speech
Kokoro-82M
by hexgrad
Kokoro‑82M is an open‑weight text‑to‑speech (TTS) model released by hexgrad . With only 82 million...
7.7Mapache-2.0
C
Text to Speech
XTTS-v2
by coqui
XTTS‑v2 is a multilingual text‑to‑speech (TTS) model released by Coqui . It builds on the original...
6.8Mmpl
Q
Text to Speech
Qwen3-TTS-12Hz-1.7B-CustomVoice
by Qwen
Qwen3‑TTS‑12Hz‑1.7B‑CustomVoice is a 1.7‑billion‑parameter text‑to‑speech (TTS) model released by...
726Kapache-2.0
R
Text to Speech
chatterbox
by ResembleAI
Model ID: ResembleAI/chatterbox | Model Name: chatterbox Chatterbox is a production‑grade,...
691Kmit

Pre-loaded AI models. Ready to run.

Skip the downloads. Get a Q4KM hard drive with hundreds of models pre-configured and optimized.

Shop Q4KM Hard Drives