openclaw - 💡(How to fix) Fix [Bug]: Telegram voice note transcription fails with Deepgram — audio arrives as raw <media:audio> without transcript [1 participants]

Summary

Telegram voice notes are no longer transcribed via the configured Deepgram provider. Deepgram transcription appears to be completely broken — it never runs.

When transcription does appear to work, it is because OpenClaw silently falls back to the local Whisper CLI (via bash). This fallback only works for agents with unrestricted exec permissions. Agents with restricted exec (e.g. exec.security: "allowlist") get zero transcription because they cannot run the Whisper fallback.

This affects all agents on the same OpenClaw instance across multiple Telegram bot accounts.

It still works fine on Discord!

Environment

• OpenClaw version: 2026.4.1
• OS: macOS 15.5 (Darwin 24.6.0, arm64, Mac Mini M4)
• Node.js: v22.22.2
• Install method: npm global
• Telegram accounts: 4 bot accounts, 2 actively tested
• Transcription provider: Deepgram nova-3 (configured explicitly)

Configuration

{
  "tools": {
    "media": {
      "audio": {
        "enabled": true,
        "language": "it",
        "providerOptions": {
          "deepgram": {
            "detect_language": false,
            "punctuate": true,
            "smart_format": true
          }
        },
        "models": [
          {
            "provider": "deepgram",
            "model": "nova-3"
          }
        ]
      }
    }
  }
}

DEEPGRAM_API_KEY is set in ~/.openclaw/.env and was working previously.

Steps to reproduce

1. Configure OpenClaw with Telegram bot + Deepgram audio transcription (as above)
2. Send a voice note to the bot via Telegram
3. Observe that Deepgram is never called — the voice note arrives as raw <media:audio> without Transcript:
4. If the agent has full exec permissions, OpenClaw may silently fall back to Whisper CLI, masking the Deepgram failure
5. If the agent has restricted exec (allowlist), no transcription occurs at all

Expected behavior

Voice notes should be transcribed via the configured Deepgram provider before being delivered to the agent.

Actual behavior

• Deepgram is never called — no Deepgram API requests are made
• OpenClaw silently falls back to Whisper CLI for agents with full exec permissions
• Agents with exec.security: "allowlist" (restricted) get no transcription at all — the Whisper fallback fails silently because bash is not in the allowlist
• No errors are logged when Deepgram transcription is skipped

Logs

Gateway error log shows a plugin ID mismatch warning (likely cosmetic but potentially related):

plugins.entries.deepgram: plugin deepgram: plugin id mismatch (manifest uses "deepgram", entry hints "deepgram-provider")

No Deepgram-related errors or API call attempts are logged. The transcription step for Deepgram appears to be silently skipped entirely.

Possibly related

• The 2026.4.1 changelog mentions: "Telegram/audio: transcode Telegram voice-note .ogg attachments before the local whisper-cli auto fallback runs, and keep mention-preflight transcription enabled in auto mode when tools.media.audio is unset"
• Issue #7899 (applyMediaUnderstanding not called)
• Issue #17101 (Telegram voice messages not transcribed)
• The Deepgram plugin ID mismatch may be causing the plugin to not be recognized as a valid transcription provider despite loading successfully

Impact

• Critical for restricted agents: agents with exec allowlist have no transcription at all — voice notes are completely unusable
• Misleading for unrestricted agents: transcription appears to work intermittently but is actually using a different provider (Whisper) than configured (Deepgram)
• Voice-heavy users locked out of voice interaction