Qwen3-TTS — Text-to-Speech API (CPU, OpenAI-compatible)

Qwen3-TTS is Alibaba's open text-to-speech model family, served here by a pure-C inference engine — no Python, no PyTorch, no GPU. Point your existing OpenAI SDK at it and get natural speech in 10 languages (EN, ZH, JA, KO, DE, FR, RU, PT, ES, IT) with 9 preset voices, per-sentence emotion markup, style instructions, and chunked audio streaming.

Endpoints

Endpoint	What
`POST /v1/audio/speech`	OpenAI-compatible: `{"input":"text","voice":"ryan"}` → WAV
`POST /v1/tts`	`{"text","speaker","emotion","instruct","volume","rate"}` → WAV
`POST /v1/tts/stream`	same body → chunked PCM stream (low time-to-first-audio)
`GET /v1/speakers`	list preset voices
`GET /v1/health`	200 when model is loaded

Inline emotion markup works in the text: "[joy] Great news! [sad] But I must go."

Quick start

curl -X POST https:///v1/tts \
  -H 'Content-Type: application/json' \
  -d '{"text":"Hello from Railway!","speaker":"ryan"}' -o hello.wav

First boot

The 1.7B model (~3.4 GB) downloads to the attached volume on first start — allow ~5–10 minutes before the service responds. The download happens once; redeploys reuse the volume.

Configuration (optional env vars)

Var	Default	Options
`MODEL`	`large`	`small` (0.6B, ~2.5× faster) · `large` (1.7B, best quality) · `voice-design` · `base-small` / `base-large`
`WORKERS`	`4`	concurrent synthesis workers
`QUANT`	`int8`	`int4` (1.7B only, least RAM) · `bf16` (full precision)
`BATCH_SIZE`	`1`	≥2 enables request batching for concurrent users

Notes

CPU-bound: synthesis is slower than realtime on shared vCPUs — use the streaming endpoint for lower perceived latency, or MODEL=small for speed.
RAM: ~4–6 GB for the default 1.7B int8 (~3 GB with MODEL=small).
Engine: bon5co/qwen3-tts (fork of gabriele-mastrapasqua/qwen3-tts with Linux server fix + container packaging). MIT / Apache-2.0 model weights.