Local speech recognition: 167 MiB of downloads and the engineering trade-offs

Local speech recognition in Loop requires about 167.3 MiB of ASR/VAD model downloads, or 189.3 MiB when expanded, plus a platform engine and temporary installation space. This implementation does not rely on cloud speech APIs; voice input and calls submit differently as the transcript goes to the configured Agent, and a phone sends audio to the Loop host. The article explains the boundary, resource pipeline, and engineering trade-offs behind local voice input.

The decision to use local inference instead of cloud-based options involves trade-offs. Local ASR avoids sending raw audio to a cloud service for text generation, reducing network usage and provider billing. However, transcripts still cross the boundary and may be sent to a cloud-hosted coding model for processing. The resource pipeline requires separate components such as VAD, ASR, TTS, and a platform engine, each with their own disk space and code costs.

The resources needed for local voice integration are downloaded, verified, and activated after the user requests installation. Installation involves more than just model files; it includes VAD, ASR, TTS, and a platform engine, with each component requiring its own versioning, verification, and repair processes. Resource management follows six constraints: explicit installation, versioning, verification, repairability, removal, and six constraints that manage these resources.