Hello Support Team, We’re seeing the assistant stop speaking on very short Japanese acknowledgements (e.g., 「はい」, 「ええ」, 「うん」) even though our stop-speaking configuration is: call id: 14d943e1-4205-4b49-87d2-e061a002a4f8 "stopSpeakingPlan": { "numWords": 0, "voiceSeconds": 0.5, "backoffSeconds": 1, "acknowledgementPhrases": ["はい"] } Issue: Utterances shorter than ~0.5s still interrupt and stop assistant speaking midway. Language: Japanese (ja) STT: Deepgram (nova-2), final transcripts show durations ~0.41s TTS: ElevenLabs Could you advise concrete parameters to prevent these micro-acknowledgements from interrupting? Thank you

To reduce interruptions from very short acknowledgements, try increasing the

voiceSeconds

parameter to its maximum recommended value. For example:

json
"stopSpeakingPlan": {
  "numWords": 0,
  "voiceSeconds": 0.5,
  "backoffSeconds": 1,
  "acknowledgementPhrases": ["はい"]
}

This will make the assistant less sensitive to brief utterances under 0.5 seconds, helping to prevent micro-acknowledgements from triggering interruptions. Test this setting and adjust slightly higher if your system allows, as per the [documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration#voice-seconds). Source: - [Voice pipeline configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration)

Since solving this issue is very important for our business, I’m sorry to bother you, but any advice you could give would be greatly appreciated.

We are reviewing the details of your request and will reach out to you when we have an update or if we need additional information.

Thank you for the support. Sorry but any update?

We have analyzed your configuration and we have some recommendations for you. Please change your configuration to this:

"stopSpeakingPlan": {
  "numWords": 2,
  "voiceSeconds": 1.0,
  "backoffSeconds": 1.5,
  "acknowledgementPhrases": [
    "はい",
    "ええ", 
    "うん",
    "そう",
    "なるほど",
    "あー",
    "うーん"
  ]
}

Also please disable smart endpointing since that is only for English.

Thank you very much. Your support is extremely helpful. However, the following error occurred: 400 - "{\"message\":[\"assistantOverrides.stopSpeakingPlan.voiceSeconds must not be greater than 0.5\"],\"error\":\"Bad Request\",\"statusCode\":400}" https://docs.vapi.ai/api-reference/calls/create#request.body.assistantOverrides.stopSpeakingPlan.numWords Also, as shown above, since numWords is not set to 0, I understand that processing is performed only based on numWords. But when using Japanese audio, if I set numWords to 2, the stopSpeakingPlan does not get triggered at all. I tried it again, and the situation is reproducible. For example, in call ID 009e1930-f921-417a-9a55-5feb18c726e1, around 0:19–0:23. If this issue can be resolved, it would greatly advance our business, so I would be very grateful for your support.

We have another configuration for you to try:

{
  "stopSpeakingPlan": {
    "numWords": 0,
    "voiceSeconds": 0.4,
    "backoffSeconds": 1.2,
    "acknowledgementPhrases": [
      "はい。",
      "ええ",
      "うん",
      "そう",
      "なるほど",
      "あー",
      "うーん"
    ],
    "interruptionPhrases": [
      "でも",
      "ちょっと",
      "すみません",
      "あの",
      "えーと",
      "いえ",
      "違います"
    ]
  }
}

Thank you, that’s very helpful. That’s exactly the kind of setting I’d like to configure, but it’s not working as expected. The assistant stops speaking in the middle and switches turns when it reacts to a “はい。” that include in the acknowledgementPhrases. call_id: 9b373098-7c39-4713-af5e-d83fa0076523 Do acknowledgementPhrases and interruptionPhrases also support Japanese? https://cdn.discordapp.com/attachments/1418151356413513788/1421359682244710431/2025-09-27_13.54.20.png?ex=68d8bfdb&is=68d76e5b&hm=8bcdd7454a964e1e11d74cc69c72218eead70214a3dd1762484e10f4802af18c&

We will need to review the logs a little more. Hang tight

Thank you for the support. Sorry but any update?

We did some analysis and have a configuration for you to try:

{
  "stopSpeakingPlan": {
    "numWords": 2,
    "voiceSeconds": 0.8,
    "backoffSeconds": 1.2,
    "acknowledgementPhrases": [
      "はい、はい。",
      "ええ、ええ。",
      "そうですね",
      "なるほど",
      "わかりました"
    ],
    "interruptionPhrases": [
      "でも",
      "ちょっと待って",
      "すみません",
      "あの、",
      "いえ、",
      "違います",
      "ちがいます"
    ]
  },
  "startSpeakingPlan": {
    "waitSeconds": 1,
    "transcriptionEndpointingPlan": {
      "onPunctuationSeconds": 1,
      "onNoPunctuationSeconds": 1,
      "onNumberSeconds": 0.5
    },
    "customEndpointingRules": [
      {
        "type": "customer",
        "regex": "^(?:はい|ええ|うん|了解|了解です|なるほど|そうですね).{0,4}[!！。．，,\\\\s]*$",
        "timeoutSeconds": 0.3
      }
    ]
  }
}

Thank you. This is exactly what I’d like to achieve if possible. However, as I mentioned in my earlier message from Sep 25, 2025, 11:42 AM, I encountered the same error again after testing: 400 - "{\"message\":[\"assistantOverrides.stopSpeakingPlan.voiceSeconds must not be greater than 0.5\"],\"error\":\"Bad Request\",\"statusCode\":400}"

Sorry for the delay. We will make a request to our team to see if they are able to change the limit

Thank you, that’s very helpful. Let me summarize the discussion in this thread. In addition to the voiceSeconds limit, could you please clarify the specifications for the following? *If these features are not currently supported, we would like to request Japanese language support, as this is extremely important for us to continue using Vapi. 1.Does the stopSpeakingPlan based on numWords support Japanese? → According to our test results, when it’s set to 2 or higher, interruptions no longer occur at all. 2.Do acknowledgementPhrases, interruptionPhrases, and customEndpointingRules support Japanese? → Based on our testing, they don’t seem to be functioning properly.

No update yet. Sorry, but we pinged the team again and will let you know once we hear back from them

Sorry but any update?

Sorry, no update yet. I'll give them a nudge to see if they can at least increase the voiceSeconds value. If there is any other properties you would like to adjust the limits to, put them here so I can ask the team for those too.

Thank you very much for your support. "increase the voiceSeconds value" will bring a dramatic improvement for users whose language does not support a numWords-based stopSpeakingPlan. - other properties you would like to adjust the limits It looks like the following items are not yet adapted for Japanese (no spaces between words, punctuation differs from English), so I’d like you to update them accordingly: For stopSpeakingPlan: numWords, acknowledgementPhrases, interruptionPhrases For startSpeakingPlan: transcriptionEndpointingPlan, customEndpointingRules Please make sure they properly support Japanese.

We are in discussions with our team to increase the

voiceSeconds

limit. For the remaining items (

acknowledgementPhrases

,

interruptionPhrases

,

numWords

, and endpointing behavior in Japanese), we’ll need to run a few tests against the current platform behavior before we implement additional updates. This will help us confirm exactly how Japanese speech patterns are being processed today and ensure we introduce changes safely. We’ll follow up as soon as we have the test results. Thanks again for your patience and for helping us improve Japanese support! ---

Thank you very much for your support. Please let me know anytime you need my (Japanese) support. If it helps speed up improvements, I genuinely want to contribute. I suspect the cause is the lack of spaces between words in Japanese. For instance, looking at the LiveKit implementation, it seems to have similar language constraints. https://github.com/livekit/agents/blob/main/livekit-agents/livekit/agents/voice/agent_activity.py#L1184

thank you for offering your support! we’ll also pass that info along to the team, and we’ll keep in touch as soon as we have updates.