Release Notes: Text To Speech#

This page tracks releases of the Text To Speech microservice. The most recent release is listed first; older entries are preserved for history.

v1.1.0#

Release Date: September 9, 2026

New:

  • Named voice selection through the voice parameter, replacing speaker-index based selection with human-readable voice identifiers.

  • Voice discovery API (GET /v1/audio/voices), including available voices and descriptions.

  • Configurable default voice support via models.tts.default_speaker.

  • Speaking-style instructions for Qwen3-TTS through the instructions request field.

Improved:

  • Synthesis performance and speech quality: faster generation with more natural prosody across supported voices.

  • OpenAI API compatibility retained for supported request fields and voice handling.

Known Issues:

  • English-only synthesis; unsupported languages return HTTP 400.

  • The model request parameter is accepted for API compatibility but the configured service model is always used.

  • Unknown voice names return HTTP 400.

v1.0.0#

Release Date: June 2026

Initial release of the Text To Speech microservice: an OpenAI-API-compatible speech synthesis service with multi-runtime support and selectable models, built for edge deployment on Intel® hardware.

New:

  • OpenAI-compatible speech endpoint (POST /v1/audio/speech) returning either raw audio/wav or a JSON envelope with metadata and a base64-encoded WAV payload.

  • Voice and model metadata endpoint (GET /v1/audio/voices) for client discovery of available speakers.

  • Multi-runtime TTS backends: openvino (Intel®-optimized) and pytorch.

  • Supported models: SpeechT5 (microsoft/speecht5_tts) and Qwen3-TTS (Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice) with custom_voice and voice_design variants.

  • Configurable device (CPU, GPU) and precision (int8, int4, fp16, fp32) where supported by the runtime/model.

  • Optional persistence of synthesized output to storage/<session_id>/ with X-Session-ID returned in the response headers.

  • Health endpoint (GET /health) for readiness probes.

  • Models are warm-loaded once per process and reused across requests to keep per-request synthesis latency low.

  • OpenVINO™ acceleration on Intel® CPUs and integrated/discrete GPUs.

  • Single config.yaml shared by standalone and container runs, with env overrides via TEXT_TO_SPEECH__....

  • Docker Compose deployment exposing the API on port 8011; standalone Python mode binds 127.0.0.1:8011 on the host.

  • Container runs as a non-root user (UID 1000).

Known issues:

  • English-only synthesis. Requests with any other language are rejected with HTTP 400.

  • The model request field is accepted for OpenAI API compatibility but is ignored; the service always uses the model defined in config.yaml.

  • For SpeechT5, language is accepted but ignored (English only). The voice field selects one of the seven bundled speaker embeddings; an unknown name returns HTTP 400.

  • Compatibility with the Video Search and Summarization sample application will be added in a subsequent release.