mtsmfm.github.io

set_asr_model

Switch how the page turns speech into text. Use it when get_status shows asr_model null (no WebGPU: the user must choose and Start stays disabled until then), when the user says transcription is lagging behind or their PC / GPU is too weak, or when they ask for better accuracy. "turbo" is whisper-large-v3-turbo in the browser (accurate, ~1.5 GB, needs a real GPU); "small" is whisper-small (~0.4 GB, runs without a GPU but is slow and makes many mistakes); "webspeech-local" is the browser's own recognizer running on the device (Chrome 139+, audio never leaves the machine, the glossary biases it); "webspeech-cloud" is the browser's own recognizer via its vendor's servers (Chrome sends the audio to Google, Edge to Microsoft; no glossary bias). The browser recognizers are free and light but do not work in Brave, Firefox or plain Chromium, and their timestamps are approximate. Ask before choosing a cloud variant — it sends the talk elsewhere. The new model is loaded right away (the status line shows the progress); switching between whisper models while recording is fine, switching between whisper and the browser recognizer takes effect at the next Start. The choice is saved in the project (project.json) and becomes the default for new projects on this device. Returns {ok, model}.

Page: /ramble-to-slides/

Input schema

{
  "type": "object",
  "properties": {
    "model": {
      "type": "string",
      "enum": [
        "turbo",
        "small",
        "webspeech-local",
        "webspeech-cloud"
      ],
      "description": "The recognizer to transcribe with from now on."
    }
  },
  "required": [
    "model"
  ]
}