Speechmatics

Cleanup Review

python scripts/_purge_cache.py is an optional cleanup helper for the Speechmatics (batch transcription) skill. It may recursively remove /tmp/skill_build/*, /tmp/skill_build/.cache. Show the exact paths first, explain whether they are skill-private temporary paths or user state, and run the cleanup only after the user confirms.

Speechmatics (batch transcription)

Transcribe an audio file via Speechmatics' async batch API. Submits a job, polls until complete, then writes the transcript.

Quick start

{baseDir}/scripts/transcribe.sh /path/to/audio.m4a

Defaults:

• Language: en
• Operating point: enhanced (better accuracy; use standard for faster/cheaper)
• Output: <input>.txt in the same directory
• Poll interval: 3s, timeout: 600s

Useful flags

{baseDir}/scripts/transcribe.sh /path/to/audio.ogg --language da
{baseDir}/scripts/transcribe.sh /path/to/meeting.wav --operating-point standard
{baseDir}/scripts/transcribe.sh /path/to/call.mp3 --diarization speaker
{baseDir}/scripts/transcribe.sh /path/to/audio.m4a --format json --out /tmp/transcript.json
{baseDir}/scripts/transcribe.sh /path/to/audio.m4a --format srt --out /tmp/subs.srt
{baseDir}/scripts/transcribe.sh /path/to/long.wav --timeout 1800

Formats: txt (default, plain text), json (Speechmatics json-v2 with word timings), srt (subtitles).

API key

The script reads the API key from (in order):

1. --api-key <key> flag
2. SPEECHMATICS_API_KEY environment variable (set by openclaw from the entry below)
3. skills.entries.speechmatics.apiKey in $OPENCLAW_CONFIG_PATH (default ~/.openclaw/openclaw.json)

Configure via openclaw.json:

{
  skills: {
    entries: {
      speechmatics: {
        apiKey: "SPEECHMATICS_KEY_HERE",
      },
    },
  },
}

Override the API base (e.g. EU region or a proxy) with --base-url or SPEECHMATICS_BASE_URL. Default: https://asr.api.speechmatics.com/v2.

Notes

• Supports most common audio formats (wav, mp3, m4a, ogg, flac, mp4, etc.) — Speechmatics transcodes server-side.
• File size limit: 2 GB per job.
• Batch jobs complete in roughly 1:10 wallclock to audio duration on enhanced; standard is faster.
• Always confirm any destructive follow-up (e.g. replying based on a transcript) before acting.