Offline Speech Service

Challenge

The video team relied on a paid third-party captioning and voiceover vendor, and that vendor was discontinued. The replacement had to run on managed enterprise hardware with no admin rights and restricted egress, meaning no cloud speech API and no new software footprint for every person who needed it. It also had to survive the kind of real-world text a video team actually produces: scripts pasted from other tools, smart quotes, mixed encodings, long-running batches that cannot be babysat.

Process

I chose Whisper for speech-to-text and Kokoro-82M via ONNX for text-to-speech, both running fully offline on the team’s existing managed hardware, with no per-user install and no dependency on an external service staying available or affordable. Getting from “it works on a clean sample” to “it works on real production scripts” took real text-handling work: cp1252 and smart-quote normalization I added so pasted-in scripts do not corrupt the pipeline, and sentence-boundary chunking so long-form narration reads naturally instead of breaking mid-thought. I put a heartbeat watchdog on the pipeline for longer batch runs so a stalled job is caught and recovered rather than left to fail silently.

Result

The service is live and running in production for the team, entirely offline and entirely within the enterprise’s IT constraints. It is internal tooling for a specific team’s workflow, not a commercial product, and it replaced a real dependency on a vendor that no longer exists with something the team fully owns and controls.