I’m seeing a recurring Polish STT issue where the transcript is missing clearly spoken words, and it can cascade into tool/logic failures because downstream decisions rely on the transcript. Example call: 019beb21-0674-7332-a511-06ef758a0474 What I observe: When listening to the call audio, I can clearly hear additional Polish words/phrases that do not appear in the transcript. Those missing words change intent/slots enough to cause the call flow to fail. Frequency: I see this fairly often (not a one-off). Environment: Mostly tested on cellular network Number imported from Twilio (Twilio → Vapi) Mitigation attempts: I tried many different settings (endpointing / thresholds / related voice-pipeline knobs) but couldn’t find a reliable mitigation. ----- "silenceTimeoutSeconds": 30, "maxDurationSeconds": 600, "endCallFunctionEnabled": True, "backgroundSound": "office", "voice": { "model": "eleven_turbo_v2_5", "voiceId": voice_id, "provider": "11labs", "stability": 0.6, "similarityBoost": 0.9, "speed": 1.1, "style": 0, "optimizeStreamingLatency": 1, "useSpeakerBoost": True, "inputMinCharacters": 3, "chunkPlan": { "enabled": True } }, "transcriber": { "provider": "deepgram", "model": "nova-3", "language": "pl", "smartFormat": True, "endpointing": 500, "confidenceThreshold": 0.1 }, "startSpeakingPlan": { "waitSeconds": 0.6, "transcriptionEndpointingPlan": { "onPunctuationSeconds": 0.6, "onNoPunctuationSeconds": 1.5, "onNumberSeconds": 1.5 } }, "stopSpeakingPlan": { "numWords": 1, "voiceSeconds": 0.3, "backoffSeconds": 1.5 }, -------

Try adding custom keyterms to your Deepgram transcriber configuration to boost recognition of critical Polish words. For example, update your config as follows:

json
"transcriber": {
  "provider": "deepgram",
  "model": "nova-3",
  "language": "pl",
  "smartFormat": true,
  "endpointing": 500,
  "confidenceThreshold": 0.1,
  "keywords": ["ważne_słowo1", "ważne_słowo2:2"]
}

Replace the keywords with the Polish words/phrases that are often missed in transcripts. This can improve transcription accuracy for those terms. For more details, see the [debugging guide](https://docs.vapi.ai/debugging). Source: - [Debugging voice agents](https://docs.vapi.ai/debugging)

Hi there, I understand you're experiencing recurring issues with Polish language transcription. Recommendation: Try Azure STT For Polish language transcription, Azure Speech-to-Text tends to perform better than other providers. It has strong support for Polish and handles various accents well. To switch to Azure STT:

{
  "transcriber": {
    "provider": "azure",
    "language": "pl-PL"
  }
}

You'll need to add your Azure Speech credentials in Dashboard > Providers. Additional tips: • Use custom keywords to boost recognition of domain-specific terms • Ensure you're using the correct language code (

pl-PL

for Polish) If you'd like me to help further, please share: 1. Call ID where the issue occurred 2. Examples of what's being transcribed incorrectly Best, Vapi Support