Trace Audio Notes & Insights

Last updated: July 26, 2026

Clearer speaker labels, without another audio upload.

After your transcript is ready, Trace uses a model bundled with the app to separate voices on this device and align them with the existing transcript. This enhancement does not rewrite what was said or upload the audio again.

  1. This device Record or import audio
  2. Google Gemini, with your consent Transcribe the audio Audio is sent for transcription
  3. This device Bundled model separates voices Audio is not uploaded again for this
  4. This device Speaker labels align to the transcript Transcript text is not rewritten
Speaker analysis runs entirely on your iPhone or iPad. Transcription and meeting notes are a separate Gemini cloud workflow that needs your consent.

Where it runsOn the current iPhone or iPad. No additional model download is needed.

What it changesSpeaker turns and labels become easier to follow. The existing transcript text remains authoritative.

If it cannot runTrace keeps the transcript and general speaker labels instead of failing the meeting.

How it works

  1. Complete the transcript first Trace first uses your selected Gemini model to create the transcript. This cloud step remains governed by the in-app data-sharing consent and Privacy Policy.
  2. Separate voices on device The bundled Core ML model analyzes the audio ranges Trace actually processed and estimates where different speakers appear. Trace does not create or retain a voiceprint that can identify someone across meetings.
  3. Align the existing content safely Trace aligns the local result with the existing transcript. A reanalysis shows a preview first, and new speaker labels are only applied after you confirm them.

Local analysis and Gemini are separate

On device

Speaker enhancement

Audio is analyzed by the bundled model on your device. This step does not send it again to Trace, FluidInference, or another service. Its output is limited to speaker ranges, labels, and alignment data for the meeting.

Google Gemini

Transcription and meeting notes

This is a separate cloud workflow. After you allow it, Trace sends the audio, transcript, meeting background, template instructions, or generated content that the requested feature needs.

In short: Speaker enhancement runs locally, but Trace's Gemini transcription and analysis features are not fully offline. See the Privacy Policy for the complete data scope.

When Trace skips it automatically

Trace skips local speaker enhancement when Low Power Mode is on, the device is running hot, the bundled model cannot load, or system resources are constrained. Your transcript continues with Gemini's general speaker labels. Once the device is ready again, you can reanalyze speakers from the result screen.

Canceling the entire processing job still stops the whole job; it is different from automatically skipping only the local speaker enhancement.

Accuracy and recovery

  • Overlapping speech, rapid interruptions, background noise, distance, and uneven volume can split one person into several speakers or merge different people.
  • Local analysis separates voices but does not know anyone's real name. You remain in control of names and merges.
  • Trace keeps the current version before applying a reanalysis preview, so you can keep the existing result or restore the previous version.
  • Reanalysis does not rewrite transcript text. It only adjusts speaker turns and timing alignment.

Open source, model, and licenses

Runtime

FluidAudio 0.15.5

Purpose
Runs speaker segmentation and clustering on Apple devices.

Bundled model

Speaker Diarization Core ML

Revision
1ed7a662fdc7109e36d822db793ee6eebdaf8594

FluidAudio also includes code derived from the VBx project and the fastcluster component. License summaries, official links, and attribution for each component remain available offline in Trace under Settings → In-App Third-Party Notices.

Need help?

If the speaker labels look incorrect, keep the current result and try reanalysis after the device cools down or Low Power Mode is turned off. See the Support page for other processing issues.