Latency measurement for Voice Translation

Latency measurement shows how long it takes for translated speech to reach the listener during Voice Translation calls. Use it to check whether delays reported by agents or customers are persistent or occasional, and to narrow down whether the network is a contributing factor.

How latency is measured

Krisp measures latency directly on the agent's computer, separately for each translation direction (agent to customer, and customer to agent). Two values are collected:

• Latency (total): the time between the moment the speaker starts talking and the moment the translated speech starts playing for the listener. This is the delay both parties experience during the call.
• Network latency: the round-trip time between the agent's computer and Krisp Cloud. It's included in the total latency, not added on top of it.

Latency is measured when the speaker starts talking after a pause and the translation of their previous speech has finished playing. When a speaker talks continuously, the listener keeps hearing translated speech without waiting, so those parts of the conversation aren't measured separately. This means not every speech item in a call has its own latency value.

Enabling latency measurement

1. Sign in to your account at account.krisp.ai
2. Click Team Settings from the left sidebar
3. Go to Products >>> Voice Translation

4. Turn on Latency measurement.
    

   

After enabling the setting, make sure agents turn off Voice Translation in the Krisp app and sync their accounts to apply the change.

Latency in Analytics

When Latency measurement is enabled, two charts are added under Analytics >>> Usage >>> Voice Translation. Both charts follow the selected date range and cadence, and can be filtered by translation direction using the Direction filter.

Voice Translation Latency over Time

This chart shows how latency changes over the selected period, helping you identify whether delays are persistent or limited to specific days.

Distribution of calls by Latency

This bar chart groups calls into three latency ranges, so you can see what share of your calls fall into each:

• Low: under 3 seconds
• Medium: 3 to 5 seconds
• High: over 5 seconds.

Latency in Call History

Latency for individual calls is available in Call History when Store call sessions and auto-check translation quality is also enabled.

Median latency per call

In Call History >>> Translations, the Median latency column shows the median latency of each call, so you can spot slower calls without opening each one. Use the Latency filter to narrow the list down.

Open a call to see its Median latency in the call details panel on the right, under Duration, Agent talked, and Noise removed.

Latency for each speech item

To see latency for each speech item in the transcript, enable Include latency information for each speech item under Store call sessions and auto-check translation quality. This option is available only when Latency measurement is turned on.

[IMAGE: Store call sessions and auto-check translation quality setting with both checkboxes visible: Include translated call recordings and Include latency information for each speech item. Highlight the latency checkbox.]

Once enabled, each speech item in the call transcript shows its latency in milliseconds under the speaker's name and timestamp:

• Hover over the value to see the breakdown of Latency and Network latency
• Values above 5 seconds are highlighted in red
• Speech items without a separate measurement show a dash (---). Hover over it to see the No separate measurement for this phrase note.

Using latency data to troubleshoot delays

Latency naturally varies within a call, so look at the median and the overall trend before focusing on individual speech items. When reviewing calls with high latency, keep in mind:

• Numbers and IDs take longer. When a speaker reads out a phone number, account number, or similar ID, Krisp waits for the complete value before translating it, so these speech items often show higher latency.
• Some languages need more context. In languages where the key part of a sentence comes at the end, such as German or Japanese, Krisp waits for more of the sentence before translating, which can raise latency on longer sentences.
• High network latency points to the connection between the agent's computer and Krisp Cloud. Review the Network requirements for Voice Translation with your IT team.