voice_assistant: option for a maximum listening time (and a graceful end of the recording)

[Nasdero]

Component name

voice_assistant

Link to component documentation on our website

https://esphome.io/components/voice_assistant/

Describe the enhancement

This discussion was drafted by Claude (an AI assistant) on behalf of Nasdero, based on measurements from his own setup. He reviewed it before it was posted.

How it works today: After the wake word, the device streams audio to Home Assistant until Home Assistant decides the speech has ended (silence detection; the device only sets VOICE_ASSISTANT_REQUEST_USE_VAD). There is no option in voice_assistant: (or in the Home Assistant Voice PE package) to limit how long the device keeps listening.

Problem: With constant background audio (radio/TV) the silence never comes. Recordings of 57 s were observed; the STT then transcribes the radio and the conversation agent can execute phantom commands.

What I'd like: a built-in option, for example max_listening_time: 6s for voice_assistant: (and ideally a number entity in the Voice PE package, so it can be changed from Home Assistant), and a defined behaviour when it triggers - ideally a graceful end of the audio stream, so that what has been recorded so far is still transcribed and processed.

Use cases

Voice PE (ESPHome 2026.10.2, stock home-assistant-voice.yaml package, Home Assistant 2026.10, local faster-whisper) in a living room with a radio/TV that is on most of the day. A hard limit of 5-7 s fits typical commands (ours take 2-4 s) and prevents long garbage recordings and wrong commands. Others run into the same: several threads on the Home Assistant forum report that the satellite "keeps listening" with background noise.

Anything else?

Workaround that works for us (template number + a script started in on_listening that calls voice_assistant.stop after N seconds, stopped again on normal end):

number:
  - platform: template
    id: max_aufnahme_s
    name: "Max. recording time"
    min_value: 3
    max_value: 15
    step: 1
    initial_value: 6
    restore_value: true
    optimistic: true

script:
  - id: recording_limit
    mode: restart
    then:
      - delay: !lambda "return (uint32_t) (id(max_aufnahme_s).state * 1000);"
      - if:
          condition:
            voice_assistant.is_running:
          then:
            - voice_assistant.stop:

voice_assistant:
  on_listening:
    - script.execute: recording_limit
  on_stt_vad_end:
    - script.stop: recording_limit
  on_stt_end:
    - script.stop: recording_limit
  on_end:
    - script.stop: recording_limit
  on_error:
    - script.stop: recording_limit
  on_idle:
    - script.stop: recording_limit

Observation / question about voice_assistant.stop: when it is called while the microphone is streaming, the pipeline run ends without stt-end (pipeline debug events: stt-start 08:52:29.8 -> run-end 08:52:34.9, no stt-end, no intent), although the STT service received the 5.05 s of audio and returned a text. The text is discarded and nothing is executed. For our use that is acceptable (it prevents phantom commands), but is this intended? Is there a way to end the audio stream gracefully so that the recording so far is still processed?

Related: the same problem is discussed on the Home Assistant forum ("Help your Assist Voice Satellite handle background noise"), where a similar timer in YAML is used as a workaround.

[github-actions[bot]]

🏷️ I've automatically added the component: voice_assistant label to help categorize this feature request.