Speech Synthesis

Overview

PCC Agent can synthesize speech on a PCC node using the TextToSpeechSE framework, letting a workflow return spoken audio to the client over the existing end-to-end-encrypted response channel rather than returning text for the device to synthesize. This supports conversational, hands-free experiences such as the PCC Agent workflows fitness coach, which can respond with synthesized audio rather than text.

Delegated tool call

Speech synthesis is implemented as a PCC Agent tool, TTSTool. A workflow invokes it on a span of text, supplying a voice, locale, and intensity, and receives synthesized audio as a stream of TTSOutput chunks.

TTSTool runs as a Delegated tool calls: a PCCAgentWorkerApp instance running on a worker node loads TextToSpeechSE and uses it in-process to synthesize the input text on the workflow’s behalf.

Voice and pronunciation asset delivery

TTSAssetManager loads the voice, locale, and pronunciation-override assets consumed by TextToSpeechSE from /private/var/MLModels/LLM at workflow initialization. They are delivered in a Codeless cryptexes packaged with the release and ratcheted into the Software Sealed Registers, so the assets are measured and included in the node’s The attestation bundle.

These assets include compiled Core ML models (.mlmodelc), a format complex enough that we don’t treat parsing it as inherently safe. As with other published codeless cryptexes, the TextToSpeechSE models are therefore held to PCC’s Verifiable Transparency guarantees and published for independent inspection using the Virtual Research Environment tools.