chatterbox-GGUF 23-Language Multilingual TTS Model: 4GB+ VRAM

chatterbox-GGUF 23-Language Multilingual TTS Model: 4GB+ VRAM

At a Glance

Item Value
Repository ggml-org/chatterbox-GGUF
Published 2026-11-06
License mit
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

ggml-org/chatterbox-GGUF, the GGUF quantized version of the open-weight multilingual text-to-speech (TTS) model “Chatterbox Multilingual V3" developed by Resemble AI, has been released by ggml-org. This model takes text as input and outputs audio corresponding to 23 languages. It has been converted to the GGUF format using ggml-org’s automatic conversion tool, enabling efficient speech generation in environments such as llama.cpp.

Specifications

The specifications and recommended settings described in the documentation for the original model “ResembleAI/chatterbox" and this GGUF distribution model are as follows:

  • Number of Parameters: 0.5B (500M)
  • Backbone Architecture: Llama backbone
  • Processing Capabilities: Text-to-speech (TTS), voice-cloning
  • Supported Languages: 23 languages (Arabic, Danish, German, Greek, English, Spanish, Finnish, French, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Dutch, Norwegian, Polish, Portuguese, Russian, Swedish, Swahili, Turkish, Chinese)
  • Training Data Volume: 0.5M hours of clean data
  • Additional Features: PerTh (Perceptual Threshold) embedded watermark output, emotion exaggeration and intensity control
  • Recommended Steps and Parameters (referencing original model):
  • General use (TTS and voice agents): Default settings of exaggeration=0.5, cfg=0.5 are suitable. If the reference speaker speaks quickly, lowering cfg to around 0.3 improves the pace. – Expressive speech: Set a lower cfg (e.g., ~0.3) and increase exaggeration to 0.7 or higher. – Controlling accent inheritance: To mitigate the retention of the reference accent when the reference audio and specified language differ, set the cfg weight to 0.
  • License Terms: Released under the MIT license, a permissive open-source license.

Performance and Quality

The model comparison table (Model Zoo) published in the model card of the original model “ResembleAI/chatterbox" is as follows:

→ Scroll horizontally to see all columns

Model Size Languages Key Features Best For Demo
Chatterbox Multilingual V3 500M 23+ Improved speaker similarity, reduced hallucinations, more natural multilingual speech Global applications, localization, cross-language voice cloning Demo
Single Language Pack 500M each 6 dedicated finetunes Language- and region-specific quality control Priority languages and dialect-sensitive applications Demos
Chatterbox 500M English CFG and exaggeration tuning General zero-shot TTS with creative controls Demo

Chatterbox Multilingual V3 supports a wide range of 23+ languages and excels in improving speaker similarity, reducing hallucinations, and generating natural multilingual speech. On the other hand, for use cases requiring strict quality control specialized for specific languages or regional dialects, the dedicated fine-tuned version Single Language Pack is more suitable. Additionally, the English-only Chatterbox excels in creative control features through CFG and exaggeration adjustments.

According to the original model card, Chatterbox Multilingual V3 exhibits higher speaker identity preservation and cross-language accent preservation compared to the conventional V2, making cross-language voice cloning more stable. It is also explained that unnecessary continuous speech and repetitive hallucinations have been reduced. It is claimed to have received high marks in comparative evaluations with commercial systems such as ElevenLabs.

Strengths and Use Cases

Based on the specifications of the original model “ResembleAI/chatterbox", this model excels in use cases that require multilingual speech synthesis and flexible voice cloning. Supported languages span 23 languages including Japanese, English, Chinese, Korean, French, German, Spanish, Russian, and Arabic, enabling reading comprehension across a wide range of languages with a single model.

The model card lists specific use cases such as video content, game audio, meme production, and interactive voice agents. By providing a reference audio file, cross-language voice cloning is possible, allowing the model to speak another language while maintaining the original speaker’s identity and accent. It is suitable for multilingual localization and global application development while maintaining the voice of the same character.

Furthermore, equipped with control functions to adjust emotional exaggeration and intensity, it is well-suited for creating dramatic and emotionally rich speech in addition to standard calm readings. The generated audio is designed to automatically embed a PerTh neural watermark that remains detectable even after audio editing and MP3 compression.

Note that when more rigorous language-specific behavior or dialect adaptation is required in Chinese, certain regional Spanish, Portuguese, or Hindi for this broad-language model (Chatterbox Multilingual V3), using the dedicated fine-tuned Single Language Pack is recommended.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 1.0GB 1.2GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as GGUF.

License — mit (Commercial use allowed): Permits commercial use, modification and redistribution, provided the copyright notice and license text are retained.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files

Weight files published in ggml-org/chatterbox-GGUF, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
chatterbox-BF16.gguf 1.05GB
chatterbox-Q8_0.gguf 559MB
mmproj-chatterbox-F16.gguf 595MB

How to Get It

This model is distributed in GGUF format via the Hugging Face ggml-org/chatterbox-GGUF repository. No prior consent request or access application is required for use.

To perform inference, a build environment of llama.cpp incorporating the necessary updates is required. Once prepared, you can execute speech output from the terminal using the llama-tts command.

llama-tts -hf ggml-org/chatterbox-GGUF -p "Hello world" --tts-lang en --output out.wav

The language to be spoken is selected with the --tts-lang option. If you want to perform voice cloning, specify a reference audio file (wav or mp3 format) using the --tts-speaker-file option. If this reference file is not specified, the default voice included with the model is used.

Additionally, you can start a local server using the llama serve command and use the text-to-speech function via the /tts endpoint.

llama serve -hf ggml-org/chatterbox-GGUF

Related Articles

What to Read Next

Sources