listen_skill

Overview

This skill improves speech-to-text accuracy over time by detecting transcription errors, learning corrections, and adapting STT behavior. It stores ultra-compact correction entries and confirms recurring patterns before applying changes automatically. The goal is fewer manual fixes and progressively cleaner transcripts.

How this skill works

The skill watches transcriptions for anomalies like wrong context or garbled names and flags likely errors. When a user corrects a transcript, the skill records the correction in a compact form and tracks occurrences. After a pattern appears two or more times, it confirms and applies the correction automatically to future transcriptions. Configuration options let you tune which STT engine to use and how aggressively to auto-apply learned fixes.

When to use it

• You have recurring transcription errors for names, acronyms, or domain terms.
• Transcripts are reviewed and corrected by humans so the system can learn.
• You want to reduce post-processing time for large batches of audio.
• You need configurable policies for when corrections should be auto-applied.

Best practices

• Make concise corrections: store mappings as compact wrong→right pairs with a confidence score.
• Require at least two independent corrections before enabling automated replacement.
• Keep a short ignore list to prevent false positives for ambiguous words.
• Periodically review learned patterns to ensure domain shifts haven’t made them obsolete.
• Tune STT configuration to match audio quality and language models before relying on auto-corrections.

Example use cases

• Correcting consistent mistranscription of a company or product name across meeting recordings.
• Training the skill to recognize industry-specific jargon that general STT models mangle.
• Reducing editor workload by auto-fixing frequent typos in a podcast transcription pipeline.
• Using conservative settings during onboarding and then increasing automation as confidence grows.