sherpa-onnx-tts

• Privacy-first voice output — you need to synthesize speech without sending text to a cloud API (e.g. sensitive documents, personal data).
• Offline or air-gapped environments — your agent runs without reliable internet access and still needs audio output.
• Cost control — you want unlimited TTS calls with no per-character or per-request billing.
• Scripted audio generation — you want to produce .wav files programmatically from a CLI command or agent tool.

• You need high-fidelity, expressive neural voices (e.g. ElevenLabs quality) — local Piper models produce good but not studio-grade results.
• You require real-time streaming TTS with ultra-low latency for interactive voice applications.
• Your platform is not x64 macOS, Linux, or Windows — no other architectures are currently bundled.

The entire TTS pipeline runs on your local machine using the sherpa-onnx native runtime. No internet connection, no API keys, and no usage fees are required after the initial model download.

Pre-built sherpa-onnx runtimes are automatically downloaded for macOS (Universal 2), Linux x64, and Windows x64, so you get the correct native binary for your OS without manual compilation.

Ships with the high-quality vits-piper-en_US-lessac-high model by default. You can drop in any other voice model from the sherpa-onnx tts-models release and point SHERPA_ONNX_MODEL_DIR at it.

Synthesize speech to a .wav file with a single command: sherpa-onnx-tts -o output.wav "Your text here". Supports overrides for model file, tokens file, and data directory via flags or environment variables.

When a model directory contains multiple .onnx files, set SHERPA_ONNX_MODEL_FILE or pass --model-file to specify which model to use, enabling multi-model setups without reconfiguring the base directory.

An AI agent summarizes a document and calls this skill to produce a spoken .wav file narration — entirely offline, with no cloud dependency in the pipeline.

A legal or medical assistant agent needs to read back sensitive content aloud. Using sherpa-onnx-tts ensures that confidential text is never transmitted to a third-party TTS API.

A developer scripts the CLI in a loop to convert dozens of text snippets into .wav files for a local application, with no per-call cost and no rate limits.

In environments where outbound internet access is blocked or restricted, this skill provides a self-contained TTS solution once the runtime and models are downloaded during setup.