Tree Inventory AI

Voice Notes

Hold-to-talk dictation per tree — Whisper transcribes your voice and Claude auto-fills structured fields.

You're standing under a tree with a clinometer in one hand and a phone in the other. Typing a paragraph of observations is the slowest part of the job. Voice notes fix that — hold the mic, talk, release. The app transcribes what you said and auto-populates the structured fields (defects, recommendations, condition) from the same recording.

Quick Start

  • Open any tree's detail screen.
  • Press and hold the microphone button.
  • Speak naturally — describe what you see.
  • Release. The transcript appears within a few seconds.
  • Claude classifies the transcript and updates the structured fields. Review and adjust.

How it works

Voice notes are a two-stage AI pipeline:

  1. OpenAI Whisper transcribes the audio into text.
  2. Claude reads the transcript and extracts structured fields — defects, recommended actions, condition rating, and free-form notes.

The original transcript is preserved verbatim. The classified fields are written into the tree record as if you had typed them by hand. Both happen within a few seconds of releasing the mic.

Recording a voice note

  1. Open the tree detail screen

    From the address view, tap any tree in the list. The detail screen shows the photo, structured fields, and a microphone button near the notes section.

  2. Press and hold the mic

    A pulsing waveform indicates the app is listening. The button stays active as long as your finger is on it.

  3. Talk naturally

    You don't need to use special syntax. Describe the tree the way you'd describe it to a colleague: "Mature red oak, around 24-inch DBH. Co-dominant stems with included bark about ten feet up. Recommend cabling the union and a Level 2 risk assessment in the next year."

  4. Release to send

    Lift your finger. The audio uploads, Whisper transcribes, Claude classifies. The transcript and structured updates land on the screen — usually within 3-6 seconds.

  5. Review the auto-fills

    Tap any field the AI updated to confirm or correct. Like photo analysis, voice classification is a starting point — your edits are the source of truth.

Multiple voice notes per tree

You can record as many voice notes per tree as you want. Each one:

  • Adds to the running list of notes on the tree
  • Re-runs classification — newer notes can override older field values if they contradict
  • Keeps its own audio file and timestamp

Use this when you walk back to the same tree after looking at it from another angle, or when a second observation comes to mind ten minutes later.

What the AI extracts

Claude is tuned to pull these fields from your transcript when you mention them:

  • Defects — co-dominant stems, included bark, deadwood, cavities, lean, soil heave, root damage, cracks, decay
  • Recommendations — pruning, removal, cabling, bracing, treatment, monitoring, Level 2 risk assessment
  • Condition rating — Good, Fair, Poor, Dead
  • Species notes — if you correct the species mid-recording ("actually that's a pin oak, not a red oak"), the species field updates
  • Free-form notes — anything that doesn't map to a structured field gets appended to the notes

If you mention something the AI doesn't recognize as a defect or recommendation, it ends up in free-form notes verbatim. Nothing is lost.

When voice notes shine

  • Hands-free in the canopy — clipboard in one hand, clinometer in the other, no free hand to type.
  • Long observations — describing crown architecture, decay patterns, or access constraints is faster spoken than typed.
  • Capturing context — easements, client preferences, neighboring property concerns, anything the structured form doesn't ask about.
  • Multiple defects in one breath — listing six things at once is much faster than tapping six checkboxes.

Quality tips

  • Distance — hold the phone roughly 30 cm (one foot) from your mouth. Closer muffles, farther drops words.
  • Wind — heavy wind degrades Whisper accuracy. Cup your hand around the mic or step behind the trunk for a windbreak.
  • Pacing — speak normally, not robotic-slow. Whisper handles natural cadence well.
  • Background noise — chainsaws, traffic, and chippers will degrade accuracy. Wait for a quiet moment if the noise is constant.
  • Pronunciation — scientific names work, but if you mumble Liriodendron tulipifera, expect a mangled transcript. Either say "tulip poplar" or speak the Latin clearly.

Privacy and data handling

  • Audio is sent to OpenAI Whisper for transcription. OpenAI does not retain audio for model training under our API plan.
  • Transcripts and classified fields are stored in your org's database, scoped to your organization only. Other orgs cannot see them.
  • The audio file persists in your org's storage so you can replay it. Deleting the voice note deletes the audio.
  • Long-term audio retention follows the same retention as the rest of your org's data — there's no separate purge schedule.

If your work involves NDAs or sensitive client information, treat voice notes the same as any other field data. Avoid speaking client names, addresses, or financial details into the recording if those don't belong in the report.

Common Questions

What if Whisper mis-transcribes a key word? Tap the transcript on the tree detail screen to edit it. After editing, tap Re-classify and Claude will re-extract structured fields from the corrected text.

Can I delete a voice note? Yes. Open the voice notes list on the tree, swipe the note left, tap delete. This removes the transcript, audio, and any field changes that note triggered.

Does the AI overwrite fields I already filled in? By default, voice classification only fills empty fields. If you've already set a defect or recommendation manually, the voice classifier won't clobber it — you'll see a "suggested change" you can accept or dismiss.

What languages does Whisper support? Whisper handles 90+ languages, but the structured-field classifier is currently tuned for English. Spanish and French transcripts will appear correctly but the field auto-fill won't run reliably yet.

Is there a length limit? Single recordings cap at about 3 minutes. If you need longer, release and start a new recording — the tree will hold multiple notes.

Related

Last updated 2026-05-03