Empty user utterance after partial Spanish-name transcription — agent stalls

Support Help
1

Hi Retell team,

We’re seeing an issue where we receive an empty user utterance during a call, which leaves our agent stalled because no follow-up event (e.g. reminder_required) fires afterward — from Retell’s perspective the user already responded, so the turn passes to us, but with empty content there’s nothing for our LLM to act on.

Context

  • The user was answering an identification question and stated their full Spanish name (first name + paternal surname + maternal surname).
  • The first name and the paternal surname were transcribed correctly.
  • When the user continued with the maternal surname, we received the following utterance on response_required:

{
“role”: “user”,
“content”: “”,
“words”: [{“word”: “”, “start”: 29.5293137, “end”: 29.5293137}]
}

content is empty, and words contains a single entry with an empty string and identical start/end timestamps. We log this as a “No content in transcript” error on our side.

Questions

  1. Is this expected in any specific scenario (low ASR confidence, partial cutoff, VAD edge case, etc.)? We’ve only seen it on Spanish-accented speech in the middle of an otherwise-correctly-transcribed utterance, which makes us suspect the transcription model (we believe English-tuned) is dropping the segment instead of producing a low-confidence transcript.
  2. Is there a way to either (a) suppress empty utterances on Retell’s side before they trigger response_required on our webhook, or (b) expose a confidence/quality signal so we can decide on our end whether to ignore the event and wait for the user to continue speaking?
  3. Can you please confirm the issue we suspect is actually the real issue?

call_id: call_d1f958098f283495ddb7e45fbed

Thanks!

2

Hello @emilia

I’ve escalated this issue to our team for further investigation.

We’ll keep you updated as soon as we have more information.

Best regards

3

Hello @emilia

  1. Switch the agent’s language setting to multilingual so the model can handle Spanish names natively. That alone should eliminate the dropped-segment behavior on accented input.

  2. Review the denoising mode — it’s set to the most aggressive option (“noise and background speech cancellation”), which can clip speech treated as non-target-language background. Stepping down to standard noise cancellation is worth testing in tandem.

  3. Defensive guard on your side: treat a response_required where content is empty (or words contains only a single zero-duration empty entry) as “keep waiting” — return an empty/no-op response and let the user keep speaking. That’s the cleanest way to avoid the stall today while the config change above addresses the underlying cause.

Thank You