Kokoro-82M, a small open-weight voice with 54 voices in 8 languages. Boubou pairs it with Qwen3.5 4B, 9B or 27B and speaks American English as af_heart.
npx boubou@beta run kokoro:82m
Talk on this Mac. It needs Node.js 18 or later; Get started walks you through.
Talk on this Linux PC. It needs an NVIDIA GPU, and the first run also downloads a 3.62 GB engine that every pipeline voice shares; Get started walks you through.
npx boubou@beta serve kokoro:82m
Talk from a phone on the same Wi-Fi, or connect a robot or app. The model still runs on this Mac. Phones and robots walks you through.
Talk from a phone on the same Wi-Fi, or connect a robot or app. The model still runs on this PC. Phones and robots walks you through.
This version runs on Apple Silicon Macs only. On Linux, choose kokoro:82m.
Apple Silicon Mac
Linux PC with an NVIDIA GPU
Memory needs are estimates. Boubou checks the machine before it downloads anything, and smaller machines still need testing.
Everything the model can speak, as its own card or files list it. Black marks what Boubou uses.
Languages
Boubou answers in American English. The card warns that support for some languages may be thin.
Source: hexgrad/Kokoro-82M
Voices
Press a voice to hear it. Each clip is the pinned model reading one sentence in that voice’s language.
American English · 11 female, 9 male
British English · 4 female, 4 male
Japanese · 4 female, 1 male
Mandarin Chinese · 4 female, 4 male
Spanish · 1 female, 2 male
French · 1 female
Hindi · 2 female, 2 male
Italian · 1 female, 1 male
Brazilian Portuguese · 1 female, 2 male
In each name, the first letter is the language and the second is f or m. Boubou downloads only af_heart.
Source: hexgrad/Kokoro-82M/voices
A version is one ready-to-run setup with its own command. Versions differ in size, language or the models they combine. Suggested ones also appear in the boubou menu.
| Version | What’s different | Speaks | Voice | Download | Mac | Linux |
|---|---|---|---|---|---|---|
| The default, with Qwen3.5-4B | American English | af_heart | 5.91 GB | 32 GB+ | NVIDIA GPU | |
| Larger language model: Qwen3.5-9B | American English | af_heart | 8.82 GB | 64 GB+ | — | |
| Largest language model: Qwen3.5-27B | American English | af_heart | 18.9 GB | 128 GB+ | — |
Silero + Smart Turn v3.2
--vad sileroParakeet TDT 0.6B v3
speech to textQwen3.5-4B · 4-bit
--llm_backend mlx-lmkokoro · af_heart
--tts kokoroBoubou owns the microphone and speakers. Apple voice processing on a Mac, or WebRTC AEC3 on Linux, cancels echo, so you can talk over a reply. Boubou drops the queued audio and cancels the generation.
Boubou maintains one Hugging Face speech-to-speech 1.1.0 pipeline and pins every model in it. Add an upstream option after the model to replace one part.
Change the language model to any MLX model on Hugging Face
npx boubou@beta run kokoro:82m --model_name mlx-community/Llama-3.2-3B-Instruct-4bit
Change speech recognition to Whisper
npx boubou@beta run kokoro:82m --stt mlx-audio-whisper
Options work with serve too. Boubou does not check the models you name: they download from Hugging Face into your own Hugging Face cache. API modules send audio or text to the server you configure. These two examples are for a Mac: on Linux the pipeline stays offline, so options must name local paths.
--vadHears when you start and stop talking. Smart Turn v3.2 then decides whether a pause ends your turn.
Not in Boubou’s runtime yet: firered
--sttWrites down what you said. Parakeet is the default; Whisper versions detect the language.
Not in Boubou’s runtime yet: faster-whisper, whisper-mlx, parakeet-unified, nemotron-streaming, orukeet, paraformer, sense-voice
--llm_backendWrites the reply. Boubou runs pinned 4-bit Qwen3.5 models with mlx-lm; --model_name picks another one.
--ttsSpeaks the reply while it is still being written. Every pipeline voice in the library is one of these.
Apache 2.0. Read the license. Weights keep their own licenses; check them before commercial use.