Text To Speech#

GitHub Readme

The Text To Speech microservice turns text into natural-sounding speech. It is designed to be dropped into voice-enabled applications (kiosks, assistants, IVR, accessibility tooling) where a simple HTTP request should return either raw WAV audio or a JSON payload with metadata.

Use Cases#

  • Voice responses for conversational assistants and kiosks.

  • Accessibility readers and announcement systems.

  • IVR / call-flow prompts generated on the edge.

  • Audio generation pipelines that need a self-hosted, OpenAI-compatible /v1/audio/speech endpoint.

Key Capabilities#

  • OpenAI-style speech endpoint and a voices/metadata endpoint.

  • Multi-runtime backends: OpenVINO (Intel-optimized) and PyTorch.

  • Configurable device (CPU, GPU) and precision (int8, int4, fp16, fp32) where the runtime/model supports it. NPU is not currently supported by any model (see Configuration > NPU).

  • Selectable speaker / voice per model family.

  • Optional persistence of synthesized output for session reuse.

Supported Models#

  • Kokoro — kokoro (default). Runs on onnxruntime CPU only; this is the model used out of the box.

  • SpeechT5 — microsoft/speecht5_tts. Lightweight, English-only, runs on CPU or GPU with models.tts.runtime: openvino. Its PyTorch implementation is not available. Does not support NPU.

  • Qwen3-TTS — Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice with model_variant: custom_voice or voice_design for richer voice control. This is the only model with NPU-specific code, but it cannot currently run on any device (CPU, GPU, or NPU) because its required qwen-tts package is not installed — see Configuration > NPU for details.

  • Runtimes: openvino (recommended on Intel hardware) and pytorch.

  • English-only synthesis in the current service build.

Next Steps#