Configure turn taking and interruptions
Turn taking has two different decisions:
- Turn completion: when Wirevox decides the caller has finished speaking and should receive a response.
- Barge-in: whether new caller speech stops audio that the agent is already playing.
A frequent mistake is tuning interruption sensitivity to fix slow responses. Use response speed or Realtime turn taking for turn completion; use the interruption controls for barge-in.
Modular controls
Response Speed
Low, Medium, and High change endpointing in the selected speech recognition provider.
- Low (~1 second) favors rapid exchanges but is most likely to commit a turn during a thinking pause.
- Medium (~3 seconds) is the balanced starting point.
- High (~5 seconds) suits deliberate explanations and long identifiers but adds latency.
These labels describe the intended experience. Each recognition provider maps them to its own endpointing or semantic turn parameters, so the results are not numerically identical across providers.
Words to Interrupt
The range is 1–5 words, with a default of 3.
- A lower value stops agent playback sooner but increases false interruptions.
- A higher value protects playback but makes the agent feel harder to interrupt.
Wirevox evaluates interim transcript length and confidence. It also suppresses probable audio echo. Do not assume that every sound immediately stops playback merely because the value is 1.
Filler Words to Ignore
This list prevents brief acknowledgements and fillers from being treated as deliberate interruptions. It is useful for words such as “mhm,” “yeah,” and “okay.” English backchannels are recognized by default, and French backchannels are added when French is a configured language.
If a caller begins with a configured filler and continues speaking, Wirevox requires additional evidence before stopping playback. A short utterance made entirely of configured fillers can be ignored as a caller turn.
Keep this list narrow. Adding meaningful words such as “cancel,” “stop,” or “wait” can make the agent talk over an important correction.
Realtime controls
Realtime providers own speech recognition and interruption detection. The Modular response-speed selector, word threshold, and filler list are not executed.
Use Voice Delivery > Turn Taking:
- Responsive commits turns sooner.
- Balanced is the default starting point.
- Patient waits longer before responding.
The selected profile maps to provider voice-activity settings. Realtime interruption events stop current playback without waiting for a Wirevox transcript word threshold.
Prompt guidance
Runtime controls decide when speech is heard and stopped. The prompt should decide how the agent behaves after an interruption:
When interrupted, stop the current explanation and address the caller's new
request directly. Do not restart or repeat the part they already heard.
Ask one focused question at a time and keep ordinary responses to one or two
short sentences unless the caller asks for more detail.
Wirevox also appends voice-channel interruption guidance at runtime. Avoid duplicating it with conflicting rules such as “always finish every sentence.”
Tune systematically
- Keep the prompt, model, voice, and test script fixed.
- Test a mid-sentence thinking pause.
- Test a brief backchannel while the agent speaks.
- Test a deliberate correction such as “No, Tuesday.”
- Test an urgent stop phrase.
- Change one timing control.
- Compare false cutoffs, failed interruptions, response delay, and transcript accuracy.
Do not optimize only for speed. The goal is the shortest delay that still preserves complete caller turns.