TTS_Paradee_Model

Paradee text to speech, an 8M-parameter model distilled from Kokoro-82M (Apache 2.0), served over HTTP so projects send text instead of loading a model themselves. It has the same admin page, tokens, rate limits, metrics, settings and update checks as DSC_Laya_Model, CNV_MiniCPM52B_Model and CNV_BitNet2B_Model. TTS_Piper_Model is the same service with the other engine, so the two can be compared on the same sentences (tools/bench.py).

The engine runs in this process on the CPU with onnxruntime. No GPU, no Ollama.

API

Call	Purpose
POST /v1/speak	{"text", "voice", "speed", "stream"} returns a WAV file (audio/wav, 16-bit mono). Headers: X-Audio-Seconds, X-Duration-Ms, X-Voice, X-Cache (hit when the kept answer was sent).
GET /v1/voices	{"voices", "default"}; with a token limited to voices, only those.
GET /health	{"status": "ok", "model", "ready", "num_thread", "voices"}; ready is true once the model is loaded. No token needed.

text is 1 to 5000 characters; speed 0.5 to 2 (default 1; above 1 is faster). Speed is passed to the model's speed input. Other fields are refused with 422, so a caller asking for mp3 finds out instead of getting WAV silently.

Streaming. With "stream": true the WAV is sent sentence by sentence while the rest is still being made, so the first audio arrives after the first sentence instead of after all of the text. The header cannot know the length yet: its size fields are 0xFFFFFFFF (read "until the connection ends"), and there is no X-Audio-Seconds. Players and ffmpeg read it; Python's wave module does not. Errors that can be known early (unknown voice, no speech in the text, model not loaded) are still normal HTTP errors; one after the first sentence cuts the stream off. Without stream the reply is the complete file with its length.

Cache. The Answers kept per replica setting (SERVICE_CACHE_SIZE, default 64, 0 = off) keeps the last answers for the same text, voice and speed; a repeat is sent again without speaking (X-Cache: hit, a few milliseconds). An answer over 1 MB (about 20 seconds of audio) is not kept, so 64 entries use at most 64 MB per replica. The cache is per replica, so with N replicas the hit rate is about 1/N, and it is emptied when the voices are reloaded (new weights or a changed thread count). A finished stream fills it too. tools/bench.py varies the speed a little so it measures speaking, not the cache.

Voices. Version 1.2 speaks af_heart (default), bf_isabella, am_puck and bm_fable from one 8.5 MB file; the British voices use British phonemes. Versions 1.1 and 1.0 (the Paradee version setting) speak af_heart only.

A token's projects list on the admin page limits the voices it may use (the page calls them projects; here they are voice names). A token without a list may use every served voice.

Admin page

/admin/ (password SERVICE_ADMIN_PASSWORD): tokens with voice scope, rate limit and expiry, usage numbers, CPU threads, the Settings card (update check hours, in-flight cap, answers kept, Paradee version, load at start), and the Updates card (Check now and Download new weights). New files are downloaded once into the shared models volume; the other replicas notice within a few seconds and load them on their next request. /metrics (admin password) is Prometheus text per token. ttsparadee_request_seconds is a histogram of how long each /v1 call took on the replica that answers (to the first byte for a streamed reply); it starts from 0 when the replica restarts.

Run

Startup, warm-up and inference never download weights. With no downloaded weights, /health answers ready: false and speech calls return 503. Open /admin/ → Updates → Download before the first call; see the admin guide for the curl alternative.

docker compose --env-file .env.compose up -d --build        # http://<host>:29321
python tools/bench.py http://127.0.0.1:29321 --save out

Without Docker: run.bat (uv, listens on 127.0.0.1:29321). Startup loads downloaded weights only; open /admin/, then Updates → Download before the first call. Memory: about 650 MB per replica, measured with both phonemizers loaded, so SERVICE_REPLICAS multiplies it. See deployment.md.

Tests

uv sync
uv run pytest

Limits worth knowing

• Speed, measured with tools/bench.py (the Piper copy) on a 12-thread laptop CPU (2026-11-09), voice af_heart: 29 to 40 times real time for sentences of 6 to 9 s of audio (0.2 s wait). Not measured on the deployment machine.

• One generation at a time per replica; extra calls wait, and more than the in-flight cap get 429.

• The rate limit is per replica.

• Output is WAV only (no ffmpeg in the image). Convert on the caller's side for MP3.

• There is no streaming: the whole clip is generated, then sent.

• The update check follows the release tag the paradee package pins (v1.2), so it reports a change only when the author moves that tag.

Try it from the admin page

Open /admin/ and use Try it to call the public API with a service token held only in memory. Runs use normal token scopes, limits and caches; the panel shows HTTP status, latency, responses, cancellation and a curl equivalent with a token placeholder. See the web UI guide.

Speak sends text, an installed voice and speed to POST /v1/speak, then displays WAV audio with playback controls and a download. Load voices with the public token: GET /v1/voices is filtered by its allowed projects list, which represents voice names for TTS. The installed Paradee version determines the available voices. An optional streaming request receives WAV chunks; the player is created when the bounded audio response finishes.

Documentation

• Purpose: input, output and what it is for
• Deployment
• Changelog
• Docker on another machine
• Docker on this PC
• Run without Docker
• Test it from the command line

Admin pages and navigation

Open /admin/ and sign in. Overview opens first; each sidebar destination has its own URL fragment and supports browser Back/Forward and direct links:

• #overview: health, model readiness, replicas, active tokens and request counts.
• #playground: service-specific request forms and streamed or media responses.
• #tokens: create, edit, expire and revoke scoped access tokens.
• #runtime: CPU threads where supported, replica information and live/restart settings.
• #updates: check versions and explicitly download weights.
• #connections: caller addresses and public endpoints.
• #models and #voices: project models and cloned voices, shown only when supported.

Collapse/Expand switches between labeled sidebar and icon rail on desktop and mobile. The non-secret sidebar preference is saved per service in this browser. Navigation preserves form drafts and active Playground requests; sign-out still clears credentials and cancels work. Page names appear once in the main header. No preview data or cross-service selector ships in the real admin UI. Existing authentication, quotas, settings precedence and explicit-download policy remain in effect.