how can i transcribe numbers or postcodes without the transcription returning too early
#how to transcribe numbers without ending early
fiery forge
pale sierraBOT
Hi @fiery forge. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:
To transcribe numbers without the transcription ending too early, you can use Deepgram's numerals feature. This feature converts spoken numbers into their numeric form in the transcript. For example, "nine two eight four seven three seven three nine two three seven" would be transcribed as "9 2 8 4 7 3 7 3 9 2 3 7". This can be particularly useful for transcribing postcodes or other numerical data.
However, if you're experiencing issues with the transcription ending too early, it might be due to the endpointing setting. The Deepgram endpointing algorithm relies on detecting silence in the audio. If you set endpointing=2000, it waits for 2000ms of silence to trigger a speech_final message. This works well with clear audio but is less effective with noisy signals, like phone calls, as it’s rare to have 2000ms of silence without some noise.
Longer endpointing times increase the likelihood of noise preventing the speech_final trigger. Shorter times (< 1000ms) work better, though it depends on the audio's noise level. By default, endpointing is set to 10ms, which is noise-resistant but may not suit cases where a clear end of speech is needed.
For times > 1000ms, the utterance_end feature is useful. It triggers based on detected silence between words in the transcript, not the audio signal. Deepgram transcriptions are produced every 3-5 seconds for is_final results, but with interim_results on, it produces transcripts every 1 second on average. So, using utterance_end with values > 1000ms can be effective.