Offline, on-device speech pipeline engineering
Offline Speech Service
Speech-to-text and text-to-speech built to run entirely on local hardware (Whisper, Kokoro), replacing a discontinued captioning and voiceover vendor.
- Role
- Sole engineer, designer, and operator
- Year
- 2025
- Status
- Live
- Whisper (speech-to-text)
- Kokoro-82M (ONNX text-to-speech)
- Python
- Managed enterprise hardware (fully offline, no cloud dependency)
Challenge
The video team’s paid captioning and voiceover vendor was discontinued. The replacement had to run on managed enterprise hardware with no admin rights and restricted egress: no cloud speech API, no per-user software installs. It also had to handle pasted scripts, smart quotes, mixed encodings, and long unattended batches.
Process
I paired Whisper (speech-to-text) with Kokoro-82M via ONNX (text-to-speech), both fully offline on the team’s existing hardware. Nothing to install per user, and no external service that could vanish or raise prices. Production scripts needed text handling: cp1252 and smart-quote normalization so pasted text doesn’t corrupt the pipeline, and sentence-boundary chunking so narration doesn’t break mid-thought. A heartbeat watchdog catches and recovers stalled batch jobs so they don’t fail silently.
Result
Deployed in production for the team, fully offline and within enterprise IT constraints. It is internal tooling for one team, not a commercial product. Both models ran on the hardware the team already had, so no audio or script was sent to an outside speech service, and the team owned its speech tooling instead of renting it from a vendor that could disappear.