Kokoro CUDA

Minimal Kokoro TTS API with CUDA support for NVIDIA GPUs (Ada Lovelace and newer).

Source: https://github.com/cipherdolls/kokoro-cuda

Features

• Single FastAPI endpoint for text-to-speech
• Streaming audio (WAV/PCM) and encoded formats (MP3, Opus, FLAC)
• 49 built-in voices (American/British English, Spanish, French, Hindi, Italian, Japanese, Portuguese, Chinese)
• Configurable voice, speed, and bitrate
• Web UI at / — type text, pick a voice, and hear audio instantly
• Swagger UI at /docs
• Automatic model + voice download on first start (persisted via Docker volumes)

Quick Start

docker build -t kokoro-cuda .
docker run --gpus all -p 8880:8880 -v kokoro-models:/app/models -v kokoro-voices:/app/voices kokoro-cuda

API

POST /v1/audio/speech

{
  "input": "Hello, this is a test.",
  "voice": "af_heart",
  "speed": 1.0,
  "response_format": "wav",
  "bitrate": "192k"
}

Parameter	Default	Options
input	(required)	Any text
voice	af_heart	49 voices — see GET /v1/voices
speed	1.0	0.5 - 2.0
response_format	wav	wav, mp3, opus, flac, pcm
bitrate	192k	128k, 192k, 320k

GET /v1/voices

Returns available voice packs.

GET /health

Returns model status.

Examples

# WAV
curl -X POST http://localhost:8880/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello world"}' -o hello.wav

# MP3 at 320k
curl -X POST http://localhost:8880/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello world", "response_format": "mp3", "bitrate": "320k"}' -o hello.mp3

Web UI

Open http://localhost:8880/ in your browser. Select a voice, adjust speed, type text, and click Speak to hear the audio immediately.

Voices

49 voice packs are downloaded automatically on first start. Naming convention: {lang}{gender}_{name}

Prefix	Language	Voices
af_ / am_	American English	20
bf_ / bm_	British English	8
ef_ / em_	Spanish	3
ff_	French	1
hf_ / hm_	Hindi	4
if_ / im_	Italian	2
jf_ / jm_	Japanese	5
pf_ / pm_	Portuguese	3
zf_	Chinese	4

Benchmark

Run the full benchmark suite (Kokoro + Whisper validation):

cd benchmark
docker compose up --build --abort-on-container-exit --exit-code-from benchmark

Results are saved to benchmark/output/<GPU_NAME>/report.md. See latest results.

Architecture

main.py            — FastAPI app, model loading, streaming TTS endpoint
download_model.py  — Downloads model + voice pack on first start
entrypoint.sh      — Runs download then starts uvicorn
Dockerfile         — CUDA 12.8 runtime image (Ada Lovelace + Blackwell)
benchmark/         — Benchmark suite with Whisper validation

Requirements

• NVIDIA GPU with CUDA 12.8+ (RTX 4090 / RTX 5090 and similar)
• Docker with NVIDIA Container Toolkit