Skip to main content

Troubleshooting a call

Look here when a call did not go as expected: a speech or model provider failed, the agent did not answer what the caller said, a reply looks glued together, the answer took long or a phrase sounds wrong. All of it shows in the call card in the full view, with Show system events switched on — the checkbox administrators and developers have.

Speech and model provider failures

When speech synthesis, speech recognition or the language model fails during a call, the card shows a TTS error, ASR error or LLM error event with a short line: provider · model · code · outcome · error text. The outcome says how the failure ended:
  • recovered by retry — a repeated request went through;
  • spoken by the fallback voice — the reply was spoken by the fallback voice instead of the main one;
  • recovered by the reserve recognition — for a recognition error: the reserve recognition carried the call on;
  • not recovered — the failure was final.
A switch of the call to a reserve is a separate event: Reserve model, Reserve recognition or Reserve voice. Its line says what replaced what and why, for example Reserve voice · switch to the reserve · Dasha → Alena · an utterance of the primary voice failed, or the primary synthesis provider is unavailable for a call that started on the reserve. Voices are labelled with their directory names. An expanded failure or switch event shows its fields as a list, the raw data is under the collapsed Event JSON. Without Show system events administrators and developers see only not recovered failures. Everything a reserve served — failures recovered through a reserve, switch events and the reserve labels on replies — and failures recovered by retry are shown in the full view. A member and a manager do not see these events; that a reserve served the call is told to them and everyone else by Reserve services were used in the header. If a reply could not be spoken by any voice, the call ends with the Platform error status: the caller would otherwise hear silence.

The caller spoke where interruption was not allowed

When a node has interruption switched off and the caller speaks over the agent’s reply anyway, their words are kept as a separate record. The conversation carries on unchanged: there was no turn for it, and the agent did not reply to it — the node did not allow itself to be interrupted. The record is visible wherever the conversation is:
  • in the call card — in both views, as an ordinary caller bubble; the said during agent’s reply label appears only in the full view;
  • in the transcript used for post-call analytics — as a Client [during agent's reply]: … line;
  • in the webhook transcript — such a line carries out_of_turn: true;
  • in report exports — as an ordinary caller utterance.
It is absent from the turn metrics (GET /v1/calls/{call_id}/turns): there was no turn, so there is nothing to measure. For the same reason it does not raise the turn count in History.
These are exactly the words you can least afford to lose. What a caller says over a disclaimer is most often “stop calling me” or “you have the wrong number”. If you build automation on the transcript — opt-outs, the do-not-call registry, complaint triage — read the records with this label too, not only the utterances inside turns.
A reply that was never heard does not count as said — and that changes the conversationA reply labelled not played goes neither into the language model’s prompt on the next turn, nor into the transcript used for post-call analytics, nor into the resolution of references like “yes, that’s him”. What the caller did not hear did not happen for the model either.Such a reply does not leave the platform either: it is not in the webhook transcript, in the call object of the Web API or in the transcript columns of report exports. It can be seen only here, in the full view, with the not played label. A reply cut off mid-way by the caller is a different case: it stays everywhere, with the interrupted mark.A lost reply does stay lost, though: the agent does not replay it. If the opening was interrupted before its first sound, the caller never hears it — and the conversation carries on as if there had been no opening.

A reply glued from fragments

A caller who pauses inside a sentence — “Yes… no, I don’t want to” — is heard by the recogniser as two turns. When the agent had not yet made a sound, the platform takes the unspoken transition back and joins the fragments into one reply. The history of such a call reads differently from what you hear on the recording, so three things are worth recognising:
  • One caller reply where the recording has two. The record carries merged_from: N — the number of fragments glued — and a fragments list with the text, timestamp and recognition phases of each. The reply’s own timestamp and its ASR and VAD phases are those of the first fragment: that is when the caller started speaking. The agent’s language model never sees the split — it is given one ordinary reply.
  • The agent appears to repeat itself. If the continuation turned out to be noise rather than speech, the transition stands and the node’s reply is simply played again, under a new turn number and with no caller reply between the two. The turn_rollback event next to it has the status redo and the reason noise. Frequent noise interruptions are a matter for the agent’s interruption settings, not for the history.
  • A transition into a node that never spoke. The rollback event’s detail carries discarded_path — the nodes the fragment took the dialogue to before it was taken back. That is the explanation for a transition in the transcript with no reply under it.
When the rollback did not happen, the same event says why with the status committed: the agent managed to speak in between (agent_spoke), the cancelled attempt had already called out to your systems (external_side_effect), the caller had heard the reply (audio_started), the four-fragment limit was reached (limit), the channel passed no sound flag (no_signal), or the rollback itself failed with an error (rollback_error) — the turn then goes on as it was. None of them affect the call’s health verdict.

Turn latency

In the full view an agent reply carries a turn N ms pill. A click opens the Turn latency card — the same one as in the test panel. A client reply has no latency line: its phases are part of the agent turn’s card. All durations are whole milliseconds. The card holds a bar of the turn made of segments, with a legend. For a voice call:
  • ASR — end of the caller’s speech → final transcript.
  • wait — final transcript → end-of-utterance commit (with streaming recognition it is the remainder of the endpointing window, with batch recognition it is near zero).
  • start — end-of-utterance commit → the agent’s reply started.
  • handoff hold — the turn waited for the continuation agent of a handoff.
  • wait check — the model decided whether the caller asks to wait.
  • Jev — the wait for the Jev transition decision, at every outcome.
  • classifier — the LLM transition classifier.
  • rewrite and retrieval — preparing the query and searching the knowledge bases.
  • LLM — the reply model only: request → first phrase handed to synthesis.
  • synth — first phrase handed to synthesis → first audio chunk.
  • playback — first chunk sent → playback started, when the platform measured it.
  • other — turn time not covered by the measured phases.
The bar shows only the segments this turn had. The segments add up to the number in the card header — the turn, from the end of the caller’s speech to the agent’s first audio. When the measured phases exceed the turn (the provider’s timings out of sync), a ⚠ sits next to the total. A text call has other segments — Jev, classifier, rewrite, retrieval and LLM — and its total is their sum. All phases opens the tree of the turn, one label per value:
  • VAD — end of speech → the decision that the utterance is over; under it, ASR and wait. When a pause split the caller’s phrase, a discarded run sits there too: the time of the dropped reply, not part of the turn sum.
  • start.
  • to TTS — end-of-utterance commit → first phrase handed to synthesis; under it, the holds, Jev with a decided by row (Jev, the LLM classifier or both), the classifier, rewrite, retrieval and LLM.
  • synth, playback, other and the turn.
  • The SDK group — the raw timings of the voice engine: LLM TTFT and TTS TTFB.
  • Mode — how text reaches synthesis: a whole phrase per request or a stream from the first token.
The hint on each row explains the phase. When the turn searched the knowledge bases, a RAG line under the card shows the number of results, the top relevance and the time. Values marked ~ are a platform estimate: the provider or the voice engine did not report exact timings for this turn.

Reply audio and the synthesis cache

Next to the ▶ button in the header of an agent reply sits the synthesis cache icon: the phrase is cached, not cached or on the denylist; the icon’s tooltip names the status, and for a reply of several phrases — “N of M phrases cached”. The ⋯ menu next to it holds the Cache actions: Re-synthesize, Delete from cache and Add to denylist. The last one asks for confirmation: the recording is removed from the cache, and the phrase is synthesized live until the mark is lifted. For a reply of several phrases the menu lists the phrases, each with the same actions and Listen to phrase.