API_KEYS: set a JSON list of keys and the API routes require one of them, in the header each client already sends: Authorization: Bearer <key> (OpenAI), xi-api-key (ElevenLabs), X-API-Key (Cartesia). /health and the web player's files stay open. Off by default. Convenience feature, be sure of your security requirements (e.g. if public, put a TLS reverse proxy in front) (#373)
ElevenLabs and Cartesia compatible websocket TTS, for voice agent frameworks (Pipecat, LiveKit Agents etc):
Streamed text input, PCM or MP3 output with word timestamps
Input with nothing to speak (..., whitespace etc) returns 0.3s of silence instead of an empty 200 or a 400, like OpenAI. UNSPOKEN_INPUT_PAUSE_S=0 for the old behavior. Websockets send it once per context, whitespace-only contexts stay empty (#579)
uvicorn (the server) bumped 0.34 to 0.54: websocket messages over 16 MB are refused, keepalive pings drop connections to clients that vanish
Model weights bake from hexgrad/Kokoro-82M at a pinned commit, checksum verified. Previous method that pinned to a GitHub release stays as a fallback. One GET per model (e.g. Inno) on HF Hub at startup, logged if the upstream configs change. See docs/configuration.md to skip.
docker/scripts/download_model.sh removed, use python docker/scripts/download_model.py --output api/src/models/v1_0
Fixed
CPU image: memory no longer keeps growing with new input lengths (jemalloc preload). Set LD_PRELOAD= to revert (#594)
CPU image RSS
v0.9.2
v0.10.0
after 100 requests
1971 MB
1552 MB
-21%
60 s idle after that
1994 MB
1318 MB
-34%
60 s after /dev/unload
1630 MB
691 MB
-58%
A download_format different from response_format is rejected with a 422 when return_download_link is set. The download link served the streamed audio unchanged under the other format's name and type, so asking for mp3 with a wav download gave an mp3 file called .wav
This PR contains the following updates:
| Package | Update | Change |
|---|---|---|
| [ghcr.io/remsky/kokoro-fastapi-cpu](https://github.com/remsky/Kokoro-FastAPI) | minor | `v0.9.2` → `v0.10.0` |
---
### Release Notes
<details>
<summary>remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-cpu)</summary>
### [`v0.10.0`](https://github.com/remsky/Kokoro-FastAPI/blob/HEAD/CHANGELOG.md#v0100---2026-11-08-WebsocketsRealtime-API-CPU-Memory--API-Keys)
[Compare Source](https://github.com/remsky/Kokoro-FastAPI/compare/v0.9.2...v0.10.0)
##### Added
- `API_KEYS`: set a JSON list of keys and the API routes require one of them, in the header each client already sends: `Authorization: Bearer <key>` (OpenAI), `xi-api-key` (ElevenLabs), `X-API-Key` (Cartesia). `/health` and the web player's files stay open. Off by default. Convenience feature, be sure of your security requirements (e.g. if public, put a TLS reverse proxy in front) ([#​373](https://github.com/remsky/Kokoro-FastAPI/issues/373))
- ElevenLabs and Cartesia compatible websocket TTS, for voice agent frameworks (Pipecat, LiveKit Agents etc):
- Streamed text input, PCM or MP3 output with word timestamps
- Supports `voice_aliases`, `lang_code`, `volume_multiplier`, `normalization_options`
- OpenAI Realtime compatible websocket TTS on `/v1/realtime` (wyoming_openai etc):
- Text input only, 24 kHz PCM output
- Supports `voice_aliases`, `lang_code`, `volume_multiplier`, `normalization_options`
##### Changed
- Input with nothing to speak (`...`, whitespace etc) returns 0.3s of silence instead of an empty 200 or a 400, like OpenAI. `UNSPOKEN_INPUT_PAUSE_S=0` for the old behavior. Websockets send it once per context, whitespace-only contexts stay empty ([#​579](https://github.com/remsky/Kokoro-FastAPI/issues/579))
- uvicorn (the server) bumped 0.34 to 0.54: websocket messages over 16 MB are refused, keepalive pings drop connections to clients that vanish
- Model weights bake from [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) at a pinned commit, checksum verified. Previous method that pinned to a GitHub release stays as a fallback. One GET per model (e.g. Inno) on HF Hub at startup, logged if the upstream configs change. See [docs/configuration.md](docs/configuration.md) to skip.
- `docker/scripts/download_model.sh` removed, use `python docker/scripts/download_model.py --output api/src/models/v1_0`
##### Fixed
- CPU image: memory no longer keeps growing with new input lengths (jemalloc preload). Set `LD_PRELOAD=` to revert ([#​594](https://github.com/remsky/Kokoro-FastAPI/issues/594))
<div align="center">
| CPU image RSS | v0.9.2 | v0.10.0 | |
| :----------------------- | ------: | ------: | ---: |
| after 100 requests | 1971 MB | 1552 MB | -21% |
| 60 s idle after that | 1994 MB | 1318 MB | -34% |
| 60 s after `/dev/unload` | 1630 MB | 691 MB | -58% |
</div>
- A `download_format` different from `response_format` is rejected with a 422 when `return_download_link` is set. The download link served the streamed audio unchanged under the other format's name and type, so asking for mp3 with a wav download gave an mp3 file called `.wav`
- OpenAI voice names added to mapped defaults:
- `"marin": "af_heart"`, `"cedar": "am_fenrir"`, `"verse": "am_puck"`, `"ballad": "bm_fable"`
</details>
---
### Configuration
📅 **Schedule**: (UTC)
- Branch creation
- At any time (no schedule defined)
- Automerge
- At any time (no schedule defined)
🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied.
♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.
🔕 **Ignore**: Close this PR and you won't be reminded about this update again.
---
- [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box
---
This PR has been generated by [Mend Renovate CLI](https://github.com/renovatebot/renovate).
<!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC4yMjIuNSIsInVwZGF0ZWRJblZlciI6IjQ0LjIyMi41IiwidGFyZ2V0QnJhbmNoIjoiZnJlc2gtc3RhcnQiLCJsYWJlbHMiOltdfQ==-->
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
This PR contains the following updates:
v0.9.2→v0.10.0Release Notes
remsky/Kokoro-FastAPI (ghcr.io/remsky/kokoro-fastapi-cpu)
v0.10.0Compare Source
Added
API_KEYS: set a JSON list of keys and the API routes require one of them, in the header each client already sends:Authorization: Bearer <key>(OpenAI),xi-api-key(ElevenLabs),X-API-Key(Cartesia)./healthand the web player's files stay open. Off by default. Convenience feature, be sure of your security requirements (e.g. if public, put a TLS reverse proxy in front) (#373)voice_aliases,lang_code,volume_multiplier,normalization_options/v1/realtime(wyoming_openai etc):voice_aliases,lang_code,volume_multiplier,normalization_optionsChanged
..., whitespace etc) returns 0.3s of silence instead of an empty 200 or a 400, like OpenAI.UNSPOKEN_INPUT_PAUSE_S=0for the old behavior. Websockets send it once per context, whitespace-only contexts stay empty (#579)docker/scripts/download_model.shremoved, usepython docker/scripts/download_model.py --output api/src/models/v1_0Fixed
LD_PRELOAD=to revert (#594)/dev/unloaddownload_formatdifferent fromresponse_formatis rejected with a 422 whenreturn_download_linkis set. The download link served the streamed audio unchanged under the other format's name and type, so asking for mp3 with a wav download gave an mp3 file called.wav"marin": "af_heart","cedar": "am_fenrir","verse": "am_puck","ballad": "bm_fable"Configuration
📅 Schedule: (UTC)
🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.
♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.
🔕 Ignore: Close this PR and you won't be reminded about this update again.
This PR has been generated by Mend Renovate CLI.