Amphion
github.com
- Free plan
- Linux
- Self-hosted
What Amphion is
Amphion is a free, open-source toolkit for generating audio, music and speech, designed to support reproducible research and help junior researchers and engineers get started. Supported tasks include text-to-speech, singing voice synthesis, voice conversion, accent conversion, singing voice conversion and text-to-audio; text-to-music is still in development. Its implementations span diffusion, transformer, VAE and flow-based architectures. Named text-to-speech models include FastSpeech2, VITS, VALL-E, NaturalSpeech2, MaskGCT and Vevo-TTS, while Vevo, FACodec and Noro are listed for voice conversion. Vevo2 supports controllable speech and singing, including editing, style conversion and melody control. Amphion also provides metrics for F0 and energy modeling, intelligibility, spectrogram distortion and speaker similarity, along with interactive model visualizations. It unifies preprocessing for multiple open-source datasets, including Emilia and the Emilia-Pipe pipeline. Installation documentation covers Python 3.9.15 with Conda or Docker; Docker setup calls for NVIDIA drivers, NVIDIA Container Toolkit and CUDA, and requires datasets to be mounted. The project is MIT-licensed and stated to be free for research and commercial use.
Who it suits
Amphion suits researchers and engineers working on audio, music or speech generation, especially those seeking a toolkit for reproducible work. It is also intended to help people starting in this area.
Good to know, and what to watch
- Supports six generation tasks, including text-to-audio.
- Includes speech, singing and voice-conversion models.
- Provides objective audio evaluation metrics.
- Unifies preprocessing for open-source datasets.
- MIT-licensed for research and commercial use.
- Text-to-music is still in development.
- Docker setup requires NVIDIA drivers, toolkit and CUDA.
- Dataset mounting is required for Docker use.
Plans
Every plan and priceAmphion
Free plan
Getting started
- Open the Amphion project on GitHub
- Prepare a Linux self-hosted environment
- Install Python 3.9.15 with Conda or use the Docker image
- For Docker, install Docker, an NVIDIA driver, NVIDIA Container Toolkit and CUDA
- Mount your dataset into the container when using Docker
- Use the supported generation or evaluation components
Limits to check
Text-to-music is listed as in development rather than supported. Docker use depends on NVIDIA drivers, NVIDIA Container Toolkit, CUDA and mounting the dataset.
The Club's review
Amphion is a free, MIT-licensed toolkit for building and studying speech, singing, and audio-generation systems. It suits researchers, engineers, and technically capable teams working on Linux or self-hosted infrastructure. Its broad model and evaluation coverage is a strong reason to adopt it; the setup burden and unfinished text-to-music work are reasons to choose something else.
Overview
Amphion combines generation models, objective evaluation, visual explanations, and dataset preparation in one open-source project. Supported work spans text-to-speech, singing voice synthesis, voice conversion, accent conversion, singing voice conversion, and text-to-audio. WAV is the export format, while audio inputs can be WAV, MP3, or FLAC.
This is a development and research environment rather than a hosted voice service. Installation is documented around Conda with Python 3.9.15 or a Docker image. The Docker route requires Docker, an NVIDIA driver, NVIDIA Container Toolkit, and CUDA, and the dataset must be mounted into the container. Teams with suitable GPU infrastructure can reproduce and extend experiments; beginners looking for an instant web workflow will face substantial preparation.
Key features
Speech and voice models
The toolkit implements a wide selection of approaches, including FastSpeech2, VITS, VALL-E, NaturalSpeech2, MaskGCT, and Vevo-TTS for text-to-speech. Vevo, FACodec, and Noro support voice-conversion work, including accent conversion. The variety is valuable when comparing architectures or building a custom pipeline, but it also means choosing and configuring models is part of the job.
Singing and controllable generation
Vevo2 extends the project into controllable speech and singing. It covers voice conversion, singing voice conversion, singing-voice editing, singing-style conversion, and melody control. That makes Amphion relevant to music-technology research as well as conventional speech synthesis, with control-oriented experiments possible inside the same toolkit.
Text-to-audio
A latent-diffusion implementation handles text-to-audio generation. It is presented as the project's implementation of the text-to-audio component of its NeurIPS 2023 work, giving developers a concrete research codebase for this task. Text-to-music, however, remains in development, so Amphion should not be selected as a ready music-generation product.
Evaluation, visualization, and data
Objective metrics cover F0 and energy modelling, intelligibility, spectrogram distortion, and speaker similarity. Interactive views explain classic processing stages, including SingVisio for the diffusion model used in singing voice conversion. Dataset preprocessing is unified across several open-source datasets and includes the Emilia dataset and Emilia-Pipe for in-the-wild speech. WeNet, Whisper, and ContentVec are among the pretrained tools used to derive content-based features.
Pricing
Amphion: free (free plan). The MIT licence permits research and commercial use, with no seat, usage, or subscription charge stated. This is the clearest fit for academic groups, independent developers, and commercial teams that can supply their own compute and engineering time. The trade-off is that free software does not remove infrastructure costs: the documented Docker stack requires NVIDIA hardware support, and you manage installation, datasets, and operation yourself.
Platforms
Amphion targets Linux and self-hosted deployments. Its documented Conda and Docker workflows are appropriate for controlled research servers or local GPU machines, not turnkey mobile or browser use.
Who it's for
The project is aimed at reproducible audio, music, and speech research and at junior researchers and engineers entering the field. It is a good choice for teams comparing diffusion, transformer, VAE, and flow-based architectures, preparing datasets, and measuring output with consistent metrics. It is a poor fit for users who need a managed API, a polished consumer interface, or supported text-to-music generation without assembling the surrounding environment.
Pros and cons
- Pro: One MIT-licensed toolkit covers speech, singing, conversion, text-to-audio, evaluation, visualization, and preprocessing, reducing the need to combine unrelated research repositories.
- Pro: The model selection supports meaningful architectural comparison, from FastSpeech2 and VITS to diffusion and flow-based systems.
- Pro: Emilia and Emilia-Pipe support in-the-wild speech preparation, useful for teams building data pipelines rather than only running demos.
- Con: Linux or self-hosting is required, and Docker adds NVIDIA drivers, NVIDIA Container Toolkit, CUDA, and dataset-mounting requirements.
- Con: Text-to-music is still in development, so the toolkit does not provide a finished workflow for that use case.
- Con: WAV export and self-managed deployment favour technical workflows over convenient, hosted sharing.
Alternatives
Choose Resemble AI instead when you want paid hosted access across web, mobile, API, and desktop environments, with a free Flex tier that uses pay-as-you-go credits that do not expire.
Altered Studio is a better fit for a commercial voice-morphing workflow with a browser or desktop interface; its free tier includes three minutes per month, 10,000 AI tokens, local voice cloning, and an attribution licence for noncommercial use.
Pick LALAL.AI for a hosted separation workflow with a free queue allowance and previews rather than a self-hosted research toolkit.
Morphic suits users who prefer managed models and workflows, with a forever-free tier capped at 20 credits for one user.
Choose Applio when you want a free, MIT-licensed open-source voice tool with Linux, macOS, Windows, web, and self-hosted options.
AudioPod AI is aimed at hosted API and web use, with a free monthly credit allowance and file-upload limits.
Pick Kits AI for a hosted generative-vocal and conversion service with a free monthly allowance and paid expansion.
Tunela is the alternative for a browser-based mix of small free allowances for conversions, stem separation, TTS, and video swapping.
For broader category comparisons, see AI Voice Conversion Software and Text-to-Speech Software.
Verdict
Choose Amphion if you need an extensible, free research foundation for speech, singing, voice conversion, text-to-audio, evaluation, and dataset work, and you can operate Linux or self-hosted GPU infrastructure. Look elsewhere if you need a managed product, broad end-user platform support, or production-ready text-to-music generation.
The verdict
Amphion combines generation models, evaluation tools and dataset preprocessing in a free research toolkit. Its documented scope does not yet include supported text-to-music generation.
Questions people ask
How much does Amphion cost?
Amphion is free under the MIT licence and is stated to be free for research and commercial use.
Which platforms does Amphion support?
The toolkit is listed for Linux and self-hosted use.
What audio formats are supported?
WAV is listed as an export format. Audio upload formats listed are WAV, MP3 and FLAC.
Can Amphion generate music from text?
Text-to-music is listed as in development. Supported tasks include text-to-speech, singing voice synthesis, voice conversion, accent conversion, singing voice conversion and text-to-audio.
What is required for Docker installation?
The Docker instructions require Docker, an NVIDIA driver, NVIDIA Container Toolkit and CUDA. They also require mounting the dataset into the container.
Is Amphion open source?
Yes. Amphion is released under the MIT License and invites contributions through its project community.
What the maker publishes
- Audio evaluation
- It provides objective evaluation metrics for F0 and energy modeling, intelligibility, spectrogram distortion, and speaker similarity. source
- Audio evaluation
- Its evaluation metrics cover F0 and energy modeling, intelligibility, spectrogram distortion, and speaker similarity. source
- Dataset preprocessing
- It unifies preprocessing for multiple open-source audio datasets and supports the Emilia dataset and Emilia-Pipe pipeline for in-the-wild speech data. source
- Datasets
- Amphion unifies preprocessing for several open-source datasets and supports the Emilia dataset and Emilia-Pipe for in-the-wild speech data. source
- Docker requirement
- The Docker instructions call for Docker, an NVIDIA Driver, NVIDIA Container Toolkit, and CUDA, and state that mounting the dataset with -v is necessary. source
- Export formats
- WAV source
- External model tools
- For content-based features, the README lists pretrained models including WeNet, Whisper, and ContentVec. source
- Free plan
- Yes source
- Generation tasks
- Supported tasks include text-to-speech, singing voice synthesis, voice conversion, accent conversion, singing voice conversion, and text-to-audio; text-to-music is listed as in development. source
- Installation
- The README documents setup using Python 3.9.15 with Conda or a Docker image, and the Docker instructions require mounting the dataset into the container. source
- Installation
- The README documents installation with a setup installer using Python 3.9.15 and with a Docker image. source
- Integrations and dependencies
- The project lists WeNet, Whisper, and ContentVec among pretrained models used for content-based features, and its Docker instructions require NVIDIA drivers, NVIDIA Container Toolkit, and CUDA. source
- Intended users
- The project describes itself as supporting reproducible research and helping junior researchers and engineers enter audio, music, and speech generation research and development. source
- License
- Amphion is released under the MIT License and is stated to be free for research and commercial use. source
- License and cost
- Amphion is licensed under MIT and the README says it is free for research and commercial use. source
- Models
- The toolkit implements diffusion, transformer, VAE, and flow-based model architectures. source
- Purpose
- Amphion is an open-source toolkit for audio, music, and speech generation aimed at supporting reproducible research and helping junior researchers and engineers get started. source
- Purpose
- Amphion is an open-source toolkit for audio, music, and speech generation intended to support reproducible research and help junior researchers and engineers get started. source
- Singing generation
- Vevo2 supports controllable speech and singing generation, including voice conversion, singing voice conversion, singing voice editing, singing style conversion, and melody control. source
- Support and community
- The project invites contributions and links to a Discord channel for community engagement. source
- Supported tasks
- The README marks text-to-speech, singing voice synthesis, voice conversion, accent conversion, singing voice conversion, and text-to-audio as supported, while text-to-music is in development. source
- Technical limit
- Text-to-music is listed as in development rather than supported. source
- Text to audio
- Amphion supports text-to-audio generation using a latent diffusion model and describes it as the official implementation of the text-to-audio generation part of its NeurIPS 2023 paper. source
- Visualization
- Amphion includes interactive visualizations of classic model processing, including SingVisio for the diffusion model used in singing voice conversion. source
Checked against the maker's pages on 5 Oct 2026.