Voice

Voice notes that transcribe while recording, then produce a summary and to-dos.

Since 0.9.1BETA

Usage

Meeting notes

Records a stand-up once in view, stops by itself after about 15 seconds, and writes a summary with to-dos; the record key or another recording hands it to you.

Loading demo...

Overview

Voice notes that transcribe while people talk: the transcript arrives speaker by speaker, and once recording stops an AI writes the summary and to-dos.

  • Once in view it presses Record: after four speakers and 2 seconds of silence it stops by itself (about 15 seconds), writes the key points and 3 to-dos, and says "3 to-dos created" inside the stage.
  • Stop works at any point while recording: only finished sentences reach the summary, and an unfinished one stays in the transcript; the key then becomes Record again.
  • Choosing another recording while one is recording or summarizing settles it at once: what was heard is summarized in place, without a toast.
  • The glow follows a simulated level and turns to its processing state while summarizing; the glow and the assistant avatar hold still when the stage is out of view, and the avatar moves only while summarizing, just after it finishes, or under the pointer.
  • Every recording plays: the audio is synthesized in the page, with no microphone and no network request, and playback starts only from your click.
  • Follow-ups answer inside the stage: Translate switches to a sample translation, while Send to Inbox and Set reminders only say they are samples; nothing is sent or scheduled.
  • The transcript and the summary follow the newest content and stop while you scroll up; the page never moves.
  • In the docs column (640–959px) the transcript sits above the summary, and from 960px the summary takes its own column with each side scrolling apart; narrow (< 640px) moves the recordings list behind a toggle in the header; reduced motion shows the finished recording with the glow off.

Customization

  • To use a real microphone, pass the stream from useMicrophone to TxVoiceBeam's stream; it analyses the level and bands itself and ignores level. Without a stream, give level a getter as this template does.
  • Append the recognizer's partial results to TxStreamText's content, keep streaming true while the person speaks, and set it to false when the sentence ends.
  • Keep the loudest level per time slice while recording and pass it as TxVoiceClip's peaks, so playback never decodes the audio; leave durationMs unset until the audio exists, because an unavailable clip holds its duration's width as a minimum and overflows a narrow column.
  • For real recordings, turn the MediaRecorder Blob into an object URL for src, and release it with URL.revokeObjectURL when the recording changes and on unmount.
  • To-do checks are stored by to-do id, not by language, so they survive switching to the translation.
RegionComponentsRole
HeaderTxTag, TxFlatRadioThe sample-data label; the list/detail toggle when narrow
RecordingsTxVoiceClipEach recording's waveform, length, and playback
RecorderTxVoiceBeamA glow that follows the level, and its processing state
TranscriptTxStreamTextSpeaker-labelled paragraphs revealed as they are spoken
SummaryTxBotAvatar, TxSkeletonThe assistant's state and the skeleton before the points
To-dos and follow-upsTxCheckbox, TxSuggestionChipsCheckable to-dos and follow-up actions
NoticeTxToastPanelThe in-stage completion notice