BestOnline.ToolsBestOnline.Tools
Voice generation

Resemble AI (Chatterbox)

Resemble AI Chatterbox: open-source speech and voice cloning

Resemble AI’s free Chatterbox models create speech and custom voices on your own hardware, with multilingual options and English Turbo and Nano models.

Open-source speechLocal generationVoice cloning

FreeAccount optional

Product previewView original
Resemble AI’s official Chatterbox Turbo artwork identifies the English model in its open-source speech family.
Resemble AI’s official Chatterbox Turbo artwork identifies the English model in its open-source speech family.Resemble AI
Access
Free
Available on
Desktop, API
OVERVIEW

Speech models you can run and adapt yourself

Resemble AI’s Chatterbox family offers self-hosted text-to-speech and reference-based voice cloning. Its August 2026 announcement ended paid voice sales to new customers while retaining support for existing accounts. Chatterbox remains open source; Resemble’s current commercial services focus on deepfake detection. A new speech project therefore starts with downloadable models rather than a new hosted voice subscription.

Multilingual V3 and the original English model have 500 million parameters. English Turbo uses 350 million, while Nano reduces this to 110 million and supports CPU inference. Six dedicated language and dialect fine-tunes cover Mandarin, Hindi, Spanish and Portuguese variants. Model files are downloaded during setup; speech generation then runs on your infrastructure, where you manage installation, integration and compute costs.

V3 must be selected explicitly: the multilingual loader still defaults to V2 when no version is supplied. Its current Python implementation accepts 23 base language codes. Czech and Vietnamese appear in the V3 release article but are absent from that accepted list. Korean is accepted, but Resemble advises against commercial V3 deployments in Korean or Vietnamese and flags limitations in Czech, Japanese and Finnish. Language-specific checkpoints and base-language coverage are separate choices.

Custom voices use reference recordings. Turbo requires a clip longer than five seconds; its example uses ten seconds. Original Chatterbox supports exaggeration and guidance controls. Turbo and Nano instead support vocal tags such as laughter and coughing, and ignore the original exaggeration and guidance adjustments. Their Python examples save WAV output. The default generation code adds a PerTh watermark, which indicates synthetic provenance without identifying a customer or model version.

Chatterbox’s MIT software licence permits commercial use and modification with the required notice. It includes no managed hosting, production support or infrastructure allowance. Permissions for a speaker’s voice and other input content remain separate. Teams looking for a ready-made browser editor or managed speech endpoint can use hosted alternatives such as ElevenLabs, Murf or Cartesia; Chatterbox suits developers who want to run and adapt the model themselves.

Best for

  • Developers building self-hosted speech applications
  • Teams needing local voice generation
  • Researchers adapting open-source voice models

Limitations

Chatterbox requires technical setup and self-managed compute. V3 must be selected explicitly; the multilingual loader defaults to V2. Czech and Vietnamese are not accepted base-language codes in the current Python implementation. Resemble advises against commercial V3 Korean/Vietnamese output. Turbo/Nano ignore original-model expression controls. Voice permissions remain separate from the MIT licence.

WHAT YOU CAN DO

Chatterbox speech-generation features

Local speech generation

Run the open-source models on your own infrastructure instead of sending each generation to a hosted voice service.

Multilingual model family

Use the general multilingual model or dedicated Mandarin, Hindi, Spanish and Portuguese checkpoints. Base language codes and regional fine-tunes have different coverage.

Reference-based custom voices

Generate speech using a supplied voice recording; Turbo’s example starts with a 10-second clip.

English Turbo and Nano

Turbo is a 350M-parameter English model; Nano uses 110M parameters and supports CPU inference. Both offer native vocal tags.

Model-specific expression

Original Chatterbox supports exaggeration and guidance. Turbo and Nano use vocal tags such as laughter and coughing, and ignore those original-model adjustments.

Open-source licensing

Use and adapt the MIT-licensed software with its required notice. Default generation adds a PerTh watermark; hosting and voice permissions are separate.

PLANS & PRICING

Resemble AI (Chatterbox) pricing

Chatterbox is free under the MIT software licence, with no included hosting or compute allowance. Your costs are hardware or rented infrastructure, integration and maintenance. Resemble ended new paid voice subscriptions in August 2026 and continues supporting existing accounts. Current detection-service plans are separate from Chatterbox speech generation.

Pricing updated 2026-10-09. Prices may change.View current pricing →
Technical specifications

Voice generation

FeatureResemble AI (Chatterbox)
Speech languages
The current Python implementation accepts 23 base-language codes; six specialised language/dialect checkpoints are separate. V3 must be selected explicitly; default loader uses V2. Czech and Vietnamese are absent from accepted base IDs. Turbo and Nano are English.Multilingual versus English model selection
Voice and pronunciation controls
Original Chatterbox offers exaggeration and guidance controls. Turbo/Nano use native paralinguistic tags such as laughter; their code ignores exaggeration and CFG adjustments.Controls differ between model architectures
Voice cloning
Reference-based cloning runs locally. Turbo requires a clip longer than 5 seconds and uses a 10-second decoder reference window; the repository example uses a 10-second clip.Self-hosted; speaker permission and audio rights required
Audio exports and streaming
The Python examples save generated speech as WAV. Default generation embeds a PerTh watermark. There is no included managed export or hosted service for new paid voice customers.Local output and self-managed deployment
Commercial-use terms
The MIT licence permits commercial use and modification of the software with its licence notice. Rights to a speaker’s voice and other input content remain separate.MIT; input rights remain the user’s responsibility
API access
SupportedSelf-hosted integration; no new paid hosted voice accounts
MORE AI TOOLS

Chatterbox alternatives

Hosted speech platforms provide managed generation and account-based usage.

Cartesia

Cartesia

Managed Sonic speech APIs with streaming, multilingual voices and paid cloning.

ElevenLabs

ElevenLabs

A hosted speech platform with a browser studio, multilingual generation and voice cloning.

Murf

Murf

A browser voiceover editor for script-based narration and production exports.

COMMON QUESTIONS

Resemble AI (Chatterbox) FAQs

Can new customers buy Resemble AI voice generation?

Resemble’s August 2026 announcement ended voice-AI sales to new customers while retaining support for existing accounts. Chatterbox remains open source and free to build on. Resemble’s current commercial focus is deepfake detection.

Is Chatterbox free for commercial projects?

The MIT software licence permits commercial use and modification with its notice. Hardware, hosting, integration and maintenance are separate costs. It does not supply permission to clone another person’s voice or use third-party content. Resemble also advises against commercial Multilingual V3 output in Korean and Vietnamese.

Does Chatterbox run locally or need a GPU?

The public Python models support self-hosted generation. Nano targets CPU and on-device use, while the original model has a Mac example using Apple’s MPS backend or CPU. Model files are downloaded during setup. Hardware needs and speed depend on the model and deployment.

Are Chatterbox Turbo and Multilingual the same model?

No. Turbo and Nano are English models with native vocal tags. The general multilingual model covers other languages, with six separate language and dialect fine-tunes. Original Chatterbox has exaggeration and guidance controls; Turbo and Nano ignore those adjustments.

Does Chatterbox load Multilingual V3 by default?

No. The current multilingual loader defaults to V2. To load V3, select t3_model="v3" when creating ChatterboxMultilingualTTS. Omitting the version, or selecting "v2", loads the older multilingual checkpoint.

Does Chatterbox support Vietnamese or Czech?

The V3 release article discusses both, but the current public Python implementation omits Czech (cs) and Vietnamese (vi) from its 23 accepted base-language codes and rejects them in language validation. Korean is accepted, although Resemble advises against commercial V3 use for Korean and Vietnamese. Japanese and Finnish also have release-specific quality limitations.

How much reference audio does Chatterbox Turbo need?

Turbo’s current implementation requires a clip longer than five seconds. The repository example uses a ten-second clip, and the decoder uses up to ten seconds of the reference. This is a model-specific input requirement; it does not describe a hosted clone-training service.

Does Chatterbox add a watermark to generated audio?

Default generation embeds Resemble’s PerTh watermark in the waveform. Python examples save the output as WAV. The V3 release describes it as a binary synthetic-provenance signal, with no customer or model-version attribution. Voice permissions remain separate.

Is this your product? Claim this page to update your listing.