Voice notes and transcription

Without accounts

Microphone audio is recorded and kept as an audio file

With your key

Deepgram or ElevenLabs (live streaming), or OpenAI (segments)

With Proofline

Transcription billed in Proofline credits (preview)

Produces

Audio, transcript.json and narration lines on the timeline

When to use it

Narrate what you expected to happen while it fails: “I removed the coupon, the total should go back to 90.” Spoken intent is the part a log cannot capture, and a developer or agent reading the report gets it at the exact moment it was said.

How it works

1. Turn on the microphone

   Use the mic control in the side panel during a recording, or set it to start with every recording. Chrome asks for microphone permission the first time.

2. Talk while you reproduce

   Audio is saved in the browser with the session.

3. Get a transcript (optional)

   In Settings → Transcription, choose a provider and enter your API key. Deepgram and ElevenLabs stream live text; OpenAI transcribes the recorded audio in segments. Each utterance is placed on the timeline at the moment you began speaking.

Providers and costs

The default models are Deepgram Nova-3, ElevenLabs Scribe v2 Realtime and OpenAI gpt-4o-transcribe, and you can change the model. Audio goes directly from the extension to the provider you chose, under your account, and that provider’s pricing and data policies apply. Without a provider, nothing is sent and the report keeps the raw audio instead of text.

With a Proofline connection (preview), transcription runs through Proofline instead and uses Proofline credits, with no provider key to manage.

Speaking annotation comments

Live annotation comments use the same recording microphone and transcription provider. With the mic already on, speech fills the open comment automatically; the comment’s mic button starts or stops the recording microphone. Deepgram and ElevenLabs show interim text as you speak, and OpenAI adds text per segment. The narration audio remains in the recording. Typing never starts the microphone.

Page vocabulary for better recognition

Product names and internal jargon are where speech recognition fails most. When you select an annotation and an AI model is configured, the extension sends a bounded excerpt of the page’s visible headings, navigation and interface labels to that model, which picks out unusual terms. A short dictionary of those terms goes to Deepgram or ElevenLabs to bias recognition.

• Input values, editable fields and hidden content are excluded from the excerpt.
• The excerpt and dictionary are kept in memory only. They are not saved with the session or published.
• No API keys are exposed to the page you are recording.
• If extraction fails, ordinary transcription continues without the dictionary.
• The model call is billed by your AI provider, or in Proofline credits when connected. ElevenLabs charges extra for keyterm prompting.

Automated tests cover this flow with mocked provider streams; recognition quality depends on the provider and has not been benchmarked here.

Limits

• Transcripts are only as good as the provider and the microphone. Proper nouns and product names may need correcting in the report.
• Narration feeds AI issue summaries as context after annotations and notes. Background chatter will be read too, so pause the mic when you are not narrating.