Dockerized FastAPI wrapper for Kokoro-82M text-to-speech model. Generate hours of high quality speech in minutes.
- OpenAI-compatible Speech + Realtime (TTS only) endpoints, multi-language support
- English (US/GB), Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, Mandarin Chinese
- ElevenLabs and Cartesia compatible websockets for voice agents (LiveKit, Pipecat)
- Custom voicepacks via Inno Clone-Tuner
- Optional integrated WebUI; read-along long-generation
- Inline multi-speaker generation, weighted voice mixing with short aliases, SSML support
- Per-word, or per-chunk timestamped caption generation
- Phoneme endpoints: generate phonemes from text, or generate audio from phonemes
- Prebuilt multiplatform images
- CPU and NVIDIA GPU (CUDA): linux/amd64 + linux/arm64
- AMD GPU (ROCm, experimental): linux/amd64 only
- Apple Silicon (MPS) supported when running directly via UV (no image)
🌐Community projects that enable Kokoro-FastAPI as a backend
- Home Assistant
- App stores and templates
- Readers and audiobooks
- Assistants and agents
- Browser
Quickest Start (docker run)
Pre-built multi-arch images with models baked in.
:latest is available, but please pin to a release tag for stable usage.
No GPU (laptop, CPU-only server)
docker run -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu:latestNVIDIA (GTX 900-series through RTX 40; ships cu126)
docker run --gpus all -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-gpu:latestNVIDIA RTX 50-series / Blackwell (ships cu128)
docker run --gpus all -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-gpu:latest-cu128NVIDIA arm64 (Jetson, GH200; same tag, ships cu129)
docker run --gpus all -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-gpu:latestAMD GPU (ROCm, experimental, x86_64 only)
docker run --device=/dev/kfd --device=/dev/dri -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-rocm:latestApple Silicon (native MPS clone; the CPU image also works)
./start-gpu_mac.shgpu:latest is the same image as gpu:latest-cu126. Configuration via environment variables, see the configuration guide.
Quick Start (docker compose)
- Install prerequisites, and start the service using Docker Compose (Full setup including UI):
- Install Docker
- Clone the repository:
git clone https://github.com/remsky/Kokoro-FastAPI.git cd Kokoro-FastAPI cd docker/gpu # For NVIDIA GPU support # or cd docker/cpu # For CPU support # or cd docker/rocm # For AMD GPU (ROCm, experimental, amd64 only) docker compose up --build # *Note for Apple Silicon (M1/M2/M3) users: # The Docker GPU image is CUDA-only and won't run on Apple Silicon. With Docker, use `docker/cpu`. # For native MPS (Apple GPU) acceleration, run directly via UV with `./start-gpu_mac.sh`. cd ../.. # back to repo root for the paths below # Models will auto-download, but if needed you can manually download: python docker/scripts/download_model.py --output api/src/models/v1_0
Configuration guide covers image vs build, the volume mounts, and env vars.
Direct Run (via uv)
- Install prerequisites:
-
Install astral-uv
-
Install espeak-ng in your system if you want it available as a fallback for unknown words/sounds. The upstream libraries may attempt to handle this, but results have varied.
-
Clone the repository:
git clone https://github.com/remsky/Kokoro-FastAPI.git cd Kokoro-FastAPIRun the model download script if you haven't already
Start directly via UV (with hot-reload)
Linux and macOS
./start-cpu.sh OR ./start-gpu.sh
Windows
.\start-cpu.ps1 OR .\start-gpu.ps1
-
Up and Running?
Run locally as an OpenAI-Compatible Speech Endpoint
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8880/v1", api_key="not-needed"
)
with client.audio.speech.with_streaming_response.create(
model="kokoro",
voice="af_sky+af_bella", #single or multiple voicepack combo
input="Hello world!"
) as response:
response.stream_to_file("output.mp3")-
The API will be available at http://localhost:8880
-
API Documentation: http://localhost:8880/docs
-
Web Interface: http://localhost:8880/web
OpenAI-Compatible Speech Endpoint
# Using OpenAI's Python library
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8880/v1", api_key="not-needed")
response = client.audio.speech.create(
model="kokoro",
voice="af_bella+af_sky", # see /api/src/core/openai_mappings.json to customize
input="Hello world!",
response_format="mp3"
)
response.stream_to_file("output.mp3")Or Via Requests:
import requests
response = requests.get("http://localhost:8880/v1/audio/voices")
voices = [v["id"] for v in response.json()["voices"]]
# Generate audio
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"model": "kokoro",
"input": "Hello world!",
"voice": "af_bella",
"response_format": "mp3", # Supported: mp3, wav, opus, flac, aac, pcm
"speed": 1.0
}
)
# Save audio
with open("output.mp3", "wb") as f:
f.write(response.content)Quick tests (run from another terminal):
python examples/assorted_checks/test_openai/test_openai_tts.py # Test OpenAI Compatibility
python examples/assorted_checks/test_voices/test_all_voices.py # Test all available voicesVoice Agents (ElevenLabs / Cartesia websocket)
ElevenLabs (multi-stream-input) and Cartesia (/tts/websocket) compatible websockets, so voice agent frameworks can use their existing plugins. Voice ids are Kokoro voice names, API keys are ignored unless API_KEYS is set.
# LiveKit Agents
from livekit.plugins import cartesia, elevenlabs
tts = elevenlabs.TTS(base_url="http://localhost:8880/v1", api_key="x", voice_id="af_heart")
tts = cartesia.TTS(base_url="http://localhost:8880", api_key="x", voice="am_michael")
# Pipecat
from pipecat.services.cartesia.tts import CartesiaTTSService
from pipecat.services.elevenlabs.tts import ElevenLabsTTSService
tts = ElevenLabsTTSService(url="ws://localhost:8880", api_key="x", voice_id="af_heart")
tts = CartesiaTTSService(url="ws://localhost:8880/tts/websocket", api_key="x", voice_id="am_michael")- Audio: PCM at 8000 to 48000 Hz, plus MP3 on the ElevenLabs route
- Word timestamps where Kokoro has them (English, espeak languages)
- Cartesia
cancelstops generation for that context. ElevenLabsclose_contextfinishes what was already sent - Speech endpoint fields go on the message that opens a context (voice tags are not supported):
voice_aliaseslang_codevolume_multipliernormalization_options
ENABLE_WEBSOCKET_TTS=falseturns these and the Realtime route off
OpenAI Realtime (TTS only)
For Realtime TTS clients (wyoming_openai TTS_REALTIME_MODELS etc). Voice names map the same as on the Speech endpoint.
import asyncio
import base64
import wave
from openai import AsyncOpenAI
async def main():
client = AsyncOpenAI(api_key="x", base_url="http://localhost:8880/v1")
audio = bytearray()
async with client.realtime.connect(model="kokoro") as conn:
await conn.session.update(session={"type": "realtime", "audio": {"output": {"voice": "af_heart"}}})
await conn.response.create(response={"conversation": "none", "input": [
{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "Hello there."}]}
]})
async for event in conn:
if event.type == "response.output_audio.delta":
audio += base64.b64decode(event.delta)
elif event.type == "response.done":
break
with wave.open("output.wav", "wb") as f:
f.setnchannels(1)
f.setsampwidth(2)
f.setframerate(24000)
f.writeframes(audio)
asyncio.run(main())- Text is read as written,
instructionsare ignored - Speech endpoint fields go in
session.updateorresponse.create:voice_aliaseslang_codevolume_multipliernormalization_options
- Audio: PCM at 24000 Hz only
response.cancelstops a response. Audio input and conversation events get anerrorback- OpenAI SDK 2.x forces
wss://, and requires you passwebsocket_base_url="ws://localhost:8880/v1"
Streaming Support
# OpenAI-compatible streaming
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8880/v1", api_key="not-needed")
# Stream to file
with client.audio.speech.with_streaming_response.create(
model="kokoro",
voice="af_bella",
input="Hello world!"
) as response:
response.stream_to_file("output.mp3")
# Stream to speakers (requires PyAudio)
import pyaudio
player = pyaudio.PyAudio().open(
format=pyaudio.paInt16,
channels=1,
rate=24000,
output=True
)
with client.audio.speech.with_streaming_response.create(
model="kokoro",
voice="af_bella",
response_format="pcm",
input="Hello world!"
) as response:
for chunk in response.iter_bytes(chunk_size=1024):
player.write(chunk)Or via requests:
import requests
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"input": "Hello world!",
"voice": "af_bella",
"response_format": "pcm"
},
stream=True
)
for chunk in response.iter_content(chunk_size=1024):
if chunk:
# Process streaming chunks
pass
Key Streaming Metrics:
- First token latency @ chunksize
- ~300ms (GPU) @ 400
- ~3500ms (CPU) @ 200 (older i7)
- ~<1s (CPU) @ 200 (M3 Pro)
- Adjustable chunking settings for real-time playback
Note: Artifacts in intonation can increase with smaller chunks
Multiple Output Audio Formats
- mp3
- wav
- opus
- flac
- aac
- pcm
Voice Combination
- Weighted voice combinations using ratios (e.g., "af_bella(2)+af_heart(1)" for 67%/33% mix)
- Ratios are automatically normalized to sum to 100%
- Available through any endpoint by adding weights in parentheses
- Saves generated voicepacks for future use
Combine voices and generate audio:
import requests
response = requests.get("http://localhost:8880/v1/audio/voices")
voices = [v["id"] for v in response.json()["voices"]]
# Weighted voice combination (67%/33% mix)
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"input": "Hello world!",
"voice": "af_bella(2)+af_sky(1)", # 2:1 ratio = 67%/33%
"response_format": "mp3"
}
)
# Download combined voice as .pt file
response = requests.post(
"http://localhost:8880/v1/audio/voices/combine",
json="af_bella(2)+af_sky(1)" # 2:1 ratio = 67%/33%
)
# Save the .pt file
with open("combined_voice.pt", "wb") as f:
f.write(response.content)
# Use the downloaded voice file
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"input": "Hello world!",
"voice": "combined_voice", # Use the saved voice file
"response_format": "mp3"
}
)
Voice Aliases
Weighted mixes can get long fast. voice_aliases maps a short name per request, for both the voice field and [voice:...] tags:
- Names match case-insensitively, but an exact match wins. Keep them lowercase to avoid surprises.
- An alias pointing at a nonexistent voice returns a 400.
- The web UI's cast exports as
{"voice_aliases": {...}}, which can be sent as-is in API calls.
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8880/v1", api_key="not-needed")
client.audio.speech.create(
model="kokoro",
voice="narrator",
input="[voice:narrator] Once upon a time. [voice:villain] Never!",
extra_body={
"allow_voice_tags": True,
"voice_aliases": {"narrator": "af_bella(2)+af_sky", "villain": "am_michael"},
},
)Voice Tuning (reference clip) 🧪
POST /dev/tune takes a 3 to 30 s clip of one English speaker and returns speech in a voice tuned toward it, via inno-kokoro. It is a tuner, not a cloner. Expect a similar voice, not a match. Only tune voices you have permission to use.
Off by default. ENABLE_INNO_TUNER=true enables the route and a Tune tab in the web player.
import json
import requests
url = "http://localhost:8880/dev/tune"
# Speak text in the tuned voice
with open("ref.wav", "rb") as ref:
response = requests.post(
url,
files={"audio": ref},
data={"request": json.dumps({"input": "Hello there."})},
)
with open("out.mp3", "wb") as f:
f.write(response.content)
# Download the tuned voice pack
with open("ref.wav", "rb") as ref:
response = requests.post(url, files={"audio": ref}, data={"return_voice_pack": "true"})
with open("ref.pt", "wb") as f:
f.write(response.content)
# Save it on the server as am_ref_tuned
with open("ref.wav", "rb") as ref:
response = requests.post(url, files={"audio": ref}, data={"save_voice": "am_ref"})
print(response.json()) # {"voice": "am_ref_tuned"}requestis the/v1/audio/speechbody withoutvoicesave_voicekeeps the pack inVOICES_DIRas<name>_tuned, usable like any other voice. Off unlessALLOW_LOCAL_VOICE_SAVING=true. Without it, the pack is deleted after the response- Saved names use the usual prefix (
am_,bf_,ax_, etc.). Bundled tuned voices end in_inno - Optional:
prosody_head(default on),fmaxpitch ceiling in Hz (default auto) - Full reference in docs/inno-tune.md
Multi-Speaker / Dialogue
[voice:...]tags switch speakers inline, anywhereinputis accepted/v1/audio/speechneedsallow_voice_tags: trueper request;/dev/dialogueallows them by defaultENABLE_VOICE_TAGS=falseopts out server-wide: the parameter is refused and/dev/dialogue403s
curl -X POST http://localhost:8880/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{
"model": "kokoro",
"voice": "af_heart",
"input": "The narrator opens. [voice:af_bella] Did it land? [pause:0.3s] [voice:am_michael] It did.",
"allow_voice_tags": true,
"response_format": "mp3"
}' --output dialogue.mp3With the official OpenAI client, pass the param in extra_body:
client.audio.speech.create(
model="kokoro",
voice="af_jadzia",
input="The narrator opens. [voice:af_bella] Did it land?",
extra_body={"allow_voice_tags": True},
)POST /dev/dialogue takes structured turns instead of tags, with an optional pause_between_turns in seconds.
import requests
response = requests.post(
"http://localhost:8880/dev/dialogue",
json={
"turns": [
{"voice": "af_bella", "text": "Did the multi speaker support land?"},
{"voice": "am_michael", "text": "It did. Turns switch voices inline."},
],
"pause_between_turns": 0.4,
"response_format": "mp3",
},
)
with open("dialogue.mp3", "wb") as f:
f.write(response.content)Notes:
- Any text before the first tag uses the request's
voice/default. - Each speaker keeps its own language pipeline based on voice prefix. An explicit
lang_codewill override every speaker. - Consecutive turns sharing a voice are merged automatically.
- Tags accept short names from Voice Aliases above, instead of full weighted mixes.
The number of voices has little impact on generation speed. Turns shorter than about 2 sentences do slow it down, from extra chunking. The cost is flat and does not grow with text length. Regenerate with examples/assorted_checks/test_dialogue/.
Inline Control Tokens
Four tokens can be embedded in the input text and are parsed server-side (API, WebUI, or any client):
- Pause:
[pause:1.5s]inserts that much silence. Must be exactly this form (colon, trailings, case-insensitive).[pause=1.5]and[PAUSE 1.0]are not recognized and get read aloud. - Pronunciation:
[Worcester](/wˈʊstər/)speaks the IPA between the slashes instead of the word. English only; use/dev/phonemizeto find the IPA. - Voice:
[voice:am_michael]switches speaker for everything that follows.- Requires
allow_voice_tags: trueper request, andENABLE_VOICE_TAGSserver-side (on by default). Otherwise the tag is spoken as written. - Accepts the same combine syntax as the
voiceparameter ([voice:af_bella(2)+af_sky]). - Short names/aliases can be defined in
voice_aliases. - Unknown values return a 400.
- Requires
- Rate:
[rate:1.5]scales the speaking voice's pace until the next rate tag or voice change;[rate:1.0]reverts. Applies on top of the requestspeed, clamped to 0.25-4.0. Same gating as voice tags.- A voice alias can carry a natural pace:
{"grandpa": {"voice": "am_michael", "rate": 0.8}}applies that rate whenever the alias speaks, as thevoiceparameter or in tags. Useful for voices that read fast or slow, and for named presets over one voice (narrator_fast,narrator_slow). - Rate belongs to the voice speaking it. A
[rate:]tag scales the speaker's own pace instead of replacing it, so an alias set to0.8stays slower under[rate:1.1]. Each[voice:...]tag resets to the new voice's own rate. Usespeedto set the pace for the whole request.
- A voice alias can carry a natural pace:
The city of [Worcester](/wˈʊstər/) is easy. [pause:1s] See?
SSML Input 🧪
Send ssml: true with allow_voice_tags: true on /v1/audio/speech or /dev/captioned_speech to translate and speak in one call. Both flags are needed, since the translation emits [voice:] and [rate:] spans that would otherwise be read aloud; ssml without them is a 400.
{
"model": "kokoro",
"input": "<speak>Hi<break time=\"750ms\"/>there</speak>",
"voice": "af_bella",
"allow_voice_tags": true,
"ssml": true
}POST /dev/ssml does the translation on its own when you want the tokens back as text rather than audio, or want to inspect them before synthesis. Send text, plus voice if your speech request uses one, then pass the result back with allow_voice_tags: true. Without a voice, <voice>/<prosody> are stripped and their content kept.
<break time="750ms"/>becomes[pause:0.75s].strength=instead oftime=gives none/x-weak 0s, weak 0.25s, medium 0.5s, strong 1s, x-strong 1.5s<voice name="am_michael">becomes[voice:am_michael], reverts at the closing tag<prosody rate="slow">becomes[rate:0.75], and takes80%or1.2too. Scales the speaking voice's pace on top of the requestspeed, clamped 0.25-4.0, reverts.pitch/volumeignored<phoneme alphabet="ipa" ph="wˈʊstər">Worcester</phoneme>becomes[Worcester](/wˈʊstər/), IPA and English only<sub alias="World Wide Web">WWW</sub>speaks the alias<desc>is dropped with its text, an audio description is not speech<emphasis>,<say-as>,<p>,<s>,<lang>, etc: markup dropped, text spoken- Malformed SSML is a 400, non-SSML passes through unchanged
- DTDs are refused and nesting past
SSML_MAX_DEPTH(10) is a 400; no dialect uses a DTD, real documents nest 2-5 - Prefixed names (
google:style,mstts:express-as,amazon:effect) need theirxmlns:on<speak>, vendor docs often omit it GET /dev/ssmlserves these tables as data, read off the translator itself
import requests
ssml = '<speak>The city of <phoneme alphabet="ipa" ph="wˈʊstər">Worcester</phoneme> is easy.<break time="1s"/>See?</speak>'
response = requests.post("http://localhost:8880/dev/ssml", json={"text": ssml})
print(response.json()) # {"text": "The city of [Worcester](/wˈʊstər/) is easy. [pause:1.0s] See?"}Natural Boundary Detection
- Automatically splits and stitches at sentence boundaries
- Reduces artifacts, and allows long-form output from a base model configured for roughly 30s at a time
The model takes up to 510 phonemized tokens per chunk, but running it that long tends to produce 'rushed' speech and other artifacts. The server adds its own chunking layer on top, sized by TARGET_MIN_TOKENS, TARGET_MAX_TOKENS, and ABSOLUTE_MAX_TOKENS (175, 250, 450 by default, set via environment variables).
Phoneme & Token Routes
Convert text to phonemes and/or generate audio directly from phonemes:
import requests
def get_phonemes(text: str, language: str = "a"):
"""Get phonemes and tokens for input text"""
response = requests.post(
"http://localhost:8880/dev/phonemize",
json={"text": text, "language": language} # "a" for American English
)
response.raise_for_status()
result = response.json()
return result["phonemes"], result["tokens"]
def generate_audio_from_phonemes(phonemes: str, voice: str = "af_bella"):
"""Generate audio from phonemes"""
response = requests.post(
"http://localhost:8880/dev/generate_from_phonemes",
json={"phonemes": phonemes, "voice": voice},
headers={"Accept": "audio/wav"}
)
if response.status_code != 200:
print(f"Error: {response.text}")
return None
return response.content
# Example usage
text = "Hello world!"
try:
# Convert text to phonemes
phonemes, tokens = get_phonemes(text)
print(f"Phonemes: {phonemes}") # e.g. ðɪs ɪz ˈoʊnli ɐ tˈɛst
print(f"Tokens: {tokens}") # Token IDs including start/end tokens
# Generate and save audio
if audio_bytes := generate_audio_from_phonemes(phonemes):
with open("speech.wav", "wb") as f:
f.write(audio_bytes)
print(f"Generated {len(audio_bytes)} bytes of audio")
except Exception as e:
print(f"Error: {e}")See examples/phoneme_examples/generate_phonemes.py for a sample script.
Timestamps (word level)
Generate audio with word-level timestamps without streaming:
import requests
import base64
import json
response = requests.post(
"http://localhost:8880/dev/captioned_speech",
json={
"model": "kokoro",
"input": "Hello world!",
"voice": "af_bella",
"speed": 1.0,
"response_format": "mp3",
"stream": False,
},
stream=False
)
with open("output.mp3","wb") as f:
audio_json=json.loads(response.content)
chunk_audio=base64.b64decode(audio_json["audio"].encode("utf-8"))
f.write(chunk_audio)
print(audio_json["timestamps"])Generate audio with word-level timestamps with streaming:
import requests
import base64
import json
response = requests.post(
"http://localhost:8880/dev/captioned_speech",
json={
"model": "kokoro",
"input": "Hello world!",
"voice": "af_bella",
"speed": 1.0,
"response_format": "mp3",
"stream": True,
},
stream=True
)
f=open("output.mp3","wb")
for chunk in response.iter_lines(decode_unicode=True):
if chunk:
chunk_json=json.loads(chunk)
chunk_audio=base64.b64decode(chunk_json["audio"].encode("utf-8"))
f.write(chunk_audio)
print(chunk_json["timestamps"])With "allow_voice_tags": true, each timestamp also carries the voice that spoke the word, so multi-speaker captions can be labelled without re-deriving the split client side. Without it the field is absent.
Timestamps (streaming chunks)
With stream, return_download_link, and return_timing set, the response carries an X-Timing-Path header pointing at a JSON sidecar of per-chunk timings. The audio body is unchanged and nothing extra is computed; this is what drives the web UI's read along.
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"input": "Hello world! [pause:1s] Again.",
"voice": "af_bella",
"stream": True,
"return_download_link": True,
"return_timing": True,
},
stream=True,
)
audio = b"".join(response.iter_content(1024))
timings = requests.get(f"http://localhost:8880/v1{response.headers['x-timing-path']}").json()
# {"chunks": [{"text": "Hello world!", "start": 0.0, "end": 0.64}, ...]}- Chunk-level (a sentence group, roughly 10-20s of audio), not word-level. For word-level use
/dev/captioned_speech. - The header arrives up front, but the file is written when generation finishes. Fetch it after the stream ends; earlier is a 404.
start/endare seconds in the final audio, so pauses and speed are already accounted for.[pause:Ns]gaps appear as{"text": ""}entries.textis normalized (numbers etc expanded), so align by words rather than exact match.- With
allow_voice_tags, each spoken chunk also carries itsvoice, same as captioned timestamps. Absent otherwise. - The sidecar sits next to the download file and shares its temp lifetime.
Performance & Benchmarks
Generation through the local API, text lengths up to feature-length books (~1.5 hours output), measuring processing time and realtime factor. Run on:
- Windows 11 Home w/ WSL2
- NVIDIA 4060Ti 16gb GPU @ CUDA 12.1
- 11th Gen i7-11700 @ 2.5GHz
- 64gb RAM
- WAV native output
- H.G. Wells - The Time Machine (full text)
Key Performance Metrics:
- Realtime Speed: Ranges between 35x-100x (generation time to output audio length)
- Average Processing Rate: 137.67 tokens/second (cl100k_base)
POST /dev/unload frees the model from VRAM and reloads lazily on the next request. The memory freed grows with load (the activation pool, not just weights), but levels off because chunks cap at 450 tokens. The long-form workload is ~30 paragraphs. Same setup as above.
| Workload | Loaded | Floor | Reclaimed | Reload |
|---|---|---|---|---|
| Short (6s audio) | 3.11 GB | 2.37 GB | 758 MiB | +4.9s |
| Long-form (7.5m) | 3.98 GB | 2.37 GB | 1,656 MiB | +5.1s |
Floor is host + CUDA context. Reproduce with uv run --extra benchmarks assorted_checks/benchmarks/benchmark_model_unload.py from examples/.
To automatically unload the model after an idle timeout, set MODEL_AUTO_UNLOAD_TIMEOUT_SECONDS to a positive number of seconds. The default 0 disables auto-unload.
End-to-end roundtrip: synthesize with Kokoro, transcribe the result back with faster-whisper, compare to the source text. Scripts and data live under examples/assorted_checks/test_transcription/.
Long-form English (full book, A Journey to the Centre of the Earth, Project Gutenberg, voice af_heart, base.en Whisper on CUDA float16, baseline captured on cu126 GPU build):
| Run | Input chars | Audio length | Synth speedup | Transcribe speedup | WER |
|---|---|---|---|---|---|
| Short (~ch.7) | 64,996 | 66m 06s | 36.4x rt | 62.4x rt | 0.047 |
| Full book | 502,766 | 507m 52s | 45.7x rt | 65.1x rt | 0.033 |
See examples/assorted_checks/test_transcription/BASELINE.md for the full regression bands.
Per-language check (single-sentence per voice, multilingual Whisper small. WER for Latin scripts, CER for ja/zh/hi):
| Language | Voice | Metric | Score |
|---|---|---|---|
| English | af_heart |
WER | 0.000 |
| English (UK) | bf_emma |
WER | 0.111 |
| Spanish | ef_dora |
WER | 0.000 |
| French | ff_siwis |
WER | 0.000 |
| Italian | if_sara |
WER | 0.000 |
| Portuguese | pf_dora |
WER | 0.000 |
| Hindi | hf_alpha |
CER | 0.059 |
| Japanese | jf_alpha |
CER | 0.000 |
| Chinese | zf_xiaobei |
CER | 0.143 |
Caveat: these are single short sentences, not a comprehensive per-language quality benchmark. They confirm each voice produces transcribable audio in its target language; deeper quality evaluation per language is open work.
To reproduce, see examples/assorted_checks/test_transcription/README.md.
Configuration Variables
Every setting is an environment variable, or a line in a .env file at the project root. Full reference in the configuration guide.
API Keys
-
A basic shared-key check, not a full auth layer. No rate limiting or scoped keys, and over HTTP keys are sent in plain text.
-
For security over a network, run the server behind a reverse proxy with HTTPS at minimum.
-
Off by default: every route stays open.
-
Set
API_KEYSto a JSON list, and API routes answer401 Unauthorizedwithout one of the keys. -
With keys set,
/health,/v1/testand/docsstay open. -
/webalso stays open currently, but the player won't load voices or generate audio.
docker run --env 'API_KEYS=["key-one","key-two"]' -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu:latestClients send it as their usual api_key:
client = OpenAI(base_url="http://localhost:8880/v1", api_key="key-one")Headers: Authorization: Bearer (OpenAI), xi-api-key (ElevenLabs), X-API-Key (Cartesia)
Debug Endpoints
System state and resource usage, for debugging exhaustion or performance issues. The /debug/* routes expose host and process internals, so they are off by default; set ENABLE_DEBUG_ENDPOINTS=true to enable.
/debug/threads- Get thread information and stack traces/debug/storage- Disk usage per mounted partition/debug/system- Get system information (CPU, memory, GPU)/dev/model- Get model load state and auto-unload timing. Off by default; setALLOW_DEV_UNLOAD=trueto enablePOST /dev/unload- Release model from VRAM; reloads lazily on next request. Off by default; setALLOW_DEV_UNLOAD=trueto enable. Returns 409 while speech is generatingPOST /dev/reload- Load the model into VRAM. Off by default; setALLOW_DEV_UNLOAD=trueto enable. Returns 409 while speech is generating
Stability: the /v1/* OpenAI-compatible routes are the stable API. /dev/* and /debug/* are operational helpers, and may change or move behind flags between minor releases.
Logging
Global API loguru logging level can be set using the API_LOG_LEVEL environment variable. Defaults to DEBUG. Per run method in the configuration guide.
Missing words & Missing some timestamps
The API normalizes input text, which can incorrectly remove or change some phrases. Disable it with "normalization_options":{"normalize": false} in the request json, the full field list is in Text normalization:
import requests
response = requests.post(
"http://localhost:8880/v1/audio/speech",
json={
"input": "Hello world!",
"voice": "af_heart",
"response_format": "pcm",
"normalization_options":
{
"normalize": False
}
},
stream=True
)
for chunk in response.iter_content(chunk_size=1024):
if chunk:
# Process streaming chunks
passLinux GPU Permissions
See docs/troubleshooting.md#linux-gpu-permissions for container group, host group, and device permission options.
AMD GPU (ROCm) troubleshooting
See docs/troubleshooting.md#amd-gpu-rocm for HSA overrides, MIOpen warmup, hipBLAS fallback, and native Linux host requirement.
WAV duration reported as nonsense in some readers
WAV responses use a streaming-sentinel (0xFFFFFFFF) for the size fields in the header. Most readers handle this fine: soundfile, pydub/ffmpeg, browsers, OS players. Python's stdlib wave does not, and reports a bogus duration. For exact length use soundfile.info(path).duration or ffprobe.
Versioning & Development
Branching Strategy:
releasebranch: Contains the latest stable build, recommended for production use. Docker images tagged with specific versions are built from this branch.masterbranch: Active development. Experimental features, ongoing changes, and fixes not yet released. Use it for the newest code, expect less stability. No images are published from here;:latestis built fromreleaselike every other tag.
Note: This is a development focused project at its core.
If something breaks, roll back to an earlier release tag, or build from source and submit a PR.
Free and open source is a community effort, and there's only really so many hours in a day. If you'd like to support the work, feel free to open a PR, buy me a coffee, or report any bugs/features/etc you find during use.
Working on the code, or pointing an AI agent at it? AGENTS.md covers the repo layout, commands, and conventions.
Model
This API uses the Kokoro-82M model from HuggingFace.
Visit the model page for more details about training, architecture, and capabilities. I have no affiliation with any of their work, and produced this wrapper for ease of use and personal projects.
License
This project is licensed under the Apache License 2.0 - see below for details:- The Kokoro model weights are licensed under Apache 2.0 (see model page)
- The FastAPI wrapper code in this repository is licensed under Apache 2.0 to match
- The inference code adapted from StyleTTS2 is MIT licensed
The full Apache 2.0 license text can be found at: https://www.apache.org/licenses/LICENSE-2.0
Project Structure and Churn
Made with contrib.rocks.
