Pronunciation Coach

Analyze spoken English pronunciation using Azure Speech Services and provide actionable coaching feedback.

Privacy Note: This skill reads local voice messages from ~/.openclaw/media/inbound/ and transmits them to Microsoft Azure Speech Services for processing.

Prerequisites

• Azure Speech API Key: Set AZURE_SPEECH_KEY env var
• Azure Speech Region: Set AZURE_SPEECH_REGION env var (e.g., southeastasia)
• ffmpeg: Required for audio format conversion (must be on PATH)
• Node.js: Required for report generation

Workflow

1. Receive Audio

Voice messages from Telegram are stored in ~/.openclaw/media/inbound/. Find the latest .ogg file matching the message timestamp.

ls -lt ~/.openclaw/media/inbound/*.ogg | head -5

2. Run Assessment

scripts/pronunciation-assess.sh <audio_file> "<reference_text>"

• audio_file: Path to the voice message (ogg/wav/mp3/m4a)
• reference_text: What the speaker intended to say (from transcript)
• The script auto-converts any format to WAV 16kHz mono

3. Generate Report

Pipe the JSON output into the report generator:

scripts/pronunciation-assess.sh audio.ogg "reference text" | node scripts/pronunciation-report.js

The report includes:

• Overall scores (Pronunciation, Accuracy, Fluency, Prosody, Completeness)
• Word-by-word breakdown with per-phoneme scores
• Problem sounds highlighted
• Verdict with actionable next steps

4. Provide Coaching

After generating the report:

1. Send the text report to the user (scores + word breakdown)
2. Identify top 3 problem sounds from the phoneme scores
3. Explain each problem — what the correct sound is and how to produce it
   • See references/phoneme-guide.md for phoneme descriptions and fixes
4. Send a voice message (via TTS) demonstrating the correct pronunciation of problem words
5. Assign practice — give the user specific sentences to re-record focusing on weak sounds

Coaching Tips

• Scores ≥ 90: Excellent, minor polish
• Scores 70-89: Good, targeted practice needed
• Scores < 70: Needs focused drill on that specific sound
• "Omission" errors mean the word wasn't detected — speaker may have been too quiet or mumbled
• Prosody score < 85 suggests monotone delivery — coach on intonation rises/falls
• Compare scores across multiple recordings to track improvement