AI voice models: Kokoro vs Chatterbox vs Fish Audio Drama 3

AI voice models, episode 5. Previous: Cartesia vs Hume. Next: Mistral Voxtral vs Microsoft.

The first four episodes covered voices you rent. This one covers voices you can download and run yourself. Hume’s voice APIs close on 13 November, as we noted last time, which is a good reminder of why some teams want weights on their own disks.

“Open” means different things here, so we picked three models that show the range. Kokoro is a tiny model with fixed voices and an Apache-2.0 licence. Chatterbox, from Resemble AI, is an MIT-licensed family that clones voices. Fish Audio Drama 3 is a 35-billion-parameter model and the highest-ranked open-weight model on the main preference board, under a licence with a revenue cap. Every figure below is from the model cards, READMEs, licence files or the Artificial Analysis text-to-speech board, read on 7 November 2026.

The three against the checklist

1. Voice cloning. Kokoro does not clone. It ships 54 preset voices, and its card describes no cloning feature. Chatterbox is built around zero-shot cloning: you pass a short reference recording (the README’s Turbo example uses a 10-second clip) and it speaks in that voice, and the repo also includes a voice conversion example. Drama 3 clones from a reference clip too; its card’s examples use a 3.9-second sample. None of the three checks consent in software. The Drama 3 licence requires prior consent to imitate an identifiable person, while Chatterbox’s card has only a one-line disclaimer.

2. Languages. Kokoro v1.0 covers 8 languages, but its voice list is uneven: 20 American and 8 British English voices, one French voice, and a warning that non-English support “may be absent or thin”. Chatterbox Multilingual V3 covers 23 languages, with separate fine-tunes for Chinese, Hindi, two Spanish and two Portuguese variants. The original Chatterbox, Turbo and Nano are English only. Drama 3’s card lists 83 language codes.

3. Latency and streaming. Kokoro is small enough that speed is rarely the issue; its card says hosted Kokoro cost under $1 per million characters as of April 2025. Resemble says Chatterbox-Nano (110M parameters) runs three times faster than real time on 8 CPU cores, and that the block-diffusion Chatterbox-Flash streams its first audio from 103 ms. Fish says Drama 3 returns first audio in a median 88.9 ms for a single stream, measured on an NVIDIA H200, and serves it through SGLang-Omni and vLLM-Omni. These are all vendor numbers on vendor hardware.

4. Emotion and style. Kokoro has a speed setting and per-voice character, not emotion control. Chatterbox has an “exaggeration” control plus a guidance setting, and Turbo and Nano accept tags such as [laugh] and [cough]. Drama 3 takes stage directions in brackets, such as [sighs], and can voice up to 16 speakers in one text.

5. Output and quality. Kokoro’s example code writes 24 kHz audio. Kokoro’s card says voices sound best between 100 and 200 tokens, are weak on very short lines and rush past about 400 tokens, so chunk long text. Resemble says Chatterbox was trained on 500,000 hours of cleaned audio. Neither Resemble’s nor Fish’s card we read states an output sample rate.

6. Price. The weights are free to download for all three; what you pay is compute. Artificial Analysis lists the hosted versions it tested at $0.65 per million characters for Kokoro, $25 for Chatterbox and $15 for Drama 3. The hardware gap is large: Kokoro is 82 million parameters, Chatterbox models run from 110 to 500 million, and Drama 3 is 35 billion with 3 billion active per token, with Fish’s speed tables all measured on H200 data-centre GPUs.

7. Speech-to-speech. None of the three is a full speech-to-speech model. Chatterbox’s voice conversion script changes the voice of existing audio. For a live agent you still need speech recognition and an LLM around any of them, the cascaded design in our voice agent pipeline explainer.

8. Long-form. Drama 3 is built for it: Fish says long text is chunked automatically and can be streamed per chunk, and earlier takes can be passed back in to edit one line. Kokoro needs you to split text yourself. Resemble says Multilingual V3 reduces “unwanted continuation, repetition, and off-prompt speech”.

9. Against the previous version. Kokoro v1.0, from January 2025, moved from under 100 hours of training audio and 10 voices to a few hundred hours and 54 voices; the Python package has not had a release since April 2025. Chatterbox has grown from one English model into a family that adds Multilingual V3, Turbo, Nano and Flash. Drama 3 is a big step up from Fish’s last open model: on the Artificial Analysis board it scores 1275 against 1108 for S2 Pro.

10. What users think. Kokoro is by far the most used: Hugging Face shows more than 12 million recent downloads, and it was among the English word error rate leaders on Hugging Face’s Open TTS Leaderboard at launch. On listener preference it is mid-table. On the Artificial Analysis board, Drama 3 is 7th overall with an Elo of 1275, the top open-weight model; Kokoro is 63rd (1064) and the original Chatterbox 85th (1023). Resemble’s own Podonos comparisons favour Chatterbox Turbo over ElevenLabs Turbo v2.5 and Cartesia Sonic 3, but those are vendor-run.

11. Licence. This is where they differ most. Kokoro is Apache-2.0, and its card welcomes deployment in real products. Chatterbox is MIT, and every file it generates carries Resemble’s Perth watermark, which Resemble says survives MP3 compression and editing. Drama 3 uses the Fish Audio Community License: free for research and for commercial use while annual revenue stays under $500,000, not as a hosted text-to-speech or cloning service for third parties at any revenue, and no removing watermarks. Our Drama 3 launch story covers the terms in full.

12. Verdict. Kokoro is the safe default for fixed voices on small hardware with no licence worries. Chatterbox is the pick if you need cloning, more languages and a permissive licence, and can live with mid-table preference scores. Drama 3 sounds the best of any open model on the board, but it needs data-centre GPUs and a licence check before any product use.

Also worth knowing

• Dia2 from Nari Labs (Apache-2.0, 1B and 2B) generates two-speaker dialogue with [S1] and [S2] tags and streams as text arrives, but only in English and up to 2 minutes.
• Sesame CSM-1B is Apache-2.0 and in Hugging Face Transformers, but the weights are gated and it needs access to Llama-3.2-1B.
• Phonon 31M from Gradium is an Apache-2.0 English model with a 64.5 MB file that runs on a CPU or in a browser. It scores 1114 on Artificial Analysis, above Kokoro. Its 40M siblings are non-commercial.
• Underdog’s Woof is even smaller, but CC BY-NC.

Why this matters

Owning the weights protects you from the kind of shutdown Hume users face this month, but only if the licence lets you ship. Read the licence file before the demo: Apache and MIT allow almost anything, while a revenue cap or a hosted-service ban can rule a model out later. For cloning, keep consent records yourself, since none of these models checks for you. Our open speech models guide covers the transcription side.

Summary

	Kokoro 82M	Chatterbox family	Fish Audio Drama 3
Size	82M	110M to 500M	35B (3B active)
Licence	Apache-2.0	MIT, watermarked output	Community License, $500K revenue cap, no hosted service
Cloning	No, 54 preset voices	Yes, short reference clip	Yes, short reference clip
Languages	8 (thin outside English)	23 (Multilingual V3)	83 language codes
Style control	Speed	Exaggeration, [laugh] tags	Bracketed directions, 16 speakers
AA Elo (7 Nov)	1064, 63rd	1023, 85th (original)	1275, 7th
Runs on	CPU or small GPU	CPU (Nano) to GPU	Data-centre GPU

We did not generate our own samples for this review. Each model card links to official demos.