Your first analysis

This page walks through the mechanics of running one analysis, end to end. For the what do the numbers mean? version, see Interpreting your first result.

Prerequisites

• Vocal Analyzer open in your browser:
  • In the browser version, open the app's web page. Nothing needs to be installed. The first analysis downloads the analysis software into the browser, which can take a minute on a slow connection; later analyses reuse it, also without a network connection.
  • In the installed app, start the local server. The easiest path is make up (Docker, all optional extras baked in) and opening http://localhost:8642. The make sync && make dev-serve path works too if you prefer uv directly; the web extra includes the WebSocket runtime needed for live microphone feedback.
• A voice recording in .wav, .mp3, .aiff, or .aif (other formats, such as .flac or .m4a, are not accepted), prepared per Recording tips.

Step 1 — upload a recording

Click Upload recordings on Home or Upload audio file in Studio, or drop a file on Home's Drop recordings to analyze area. In the installed app the upload limit is 500 MB; most speech recordings are well under 5 MB.

A progress bar shows the status:

• Queued for analysis — waiting for an analysis worker.
• Loading audio…, Extracting pitch, formants, HNR…, Building timeline data… — pipeline progress.
• Transcribing with Whisper… — optional transcription step in the installed app, if Whisper is installed. The browser version does not transcribe.
• Complete — results are ready.

For short recordings this finishes in a few seconds. In the installed app, long recordings with transcription can take several minutes, especially on CPU-only hardware.

Note: In the installed app, if the job stays in one progress state for more than 30 minutes, the server marks it as error with a message suggesting a shorter recording or a smaller Whisper model. The dashboard surfaces that message directly.

In the browser version, the analysis runs in the browser tab. Keep the tab open until it finishes. If you close the tab during an analysis, the recording is kept and appears as a failed analysis with a Retry button the next time you open the app.

Step 2 — scan the dashboard

Once the job finishes, you should see:

• A player bar with play/pause, a time readout, and a playback-speed toggle.
• The timeline canvas with the waveform, pitch track (F0), and formant tracks (F1/F2) plotted together.
• The summary panel on the right with:
  • The gender-perception spectrum bars for the primary cues (pitch and formant dispersion).
  • The voice detail tile grid for secondary metrics (HNR, voice quality, LTAS complements, sibilants, etc.).
  • Vowel space and per-vowel resonance charts when the recording contains enough distinct vowel classes.
  • Detailed measurements — raw frequency rows for F1/F2/F3, ceiling, voiced frame count, duration, and sample rate.
• The history sidebar on the left, where the analysis now appears as a saved session.

Step 3 — read the result

Three things to check before diving into specific numbers:

1. Measurement warnings. Red or yellow warning tiles at the top of the summary panel tell you when a measurement is less reliable than usual (low SNR, sample rate below 16 kHz, F0 close to F1, etc.). These are informational, not blocking — but a warning means that metric's displayed value should be interpreted cautiously.
2. Confidence badges. Secondary tiles display a small confidence label (high / moderate / low / unavailable). A metric at low or unavailable confidence has either degraded extraction quality or no usable data for this recording.
3. Spectrum bar positions. The two primary cues — pitch and formant dispersion — sit as diamonds on a three-zone bar. That position is the fastest read of where the recording lands relative to published central tendencies.

Step 4 — compare, if useful

Every analysis shows up as a session in the left sidebar. Click an older session to load it back into the main view and compare with the current one. Re-running the same audio with different settings (via Rerun in the session menu) does not delete the original.

Caveat: Benchmark continuity is anchored to the whole-recording formant mode. If you change the formant mode between sessions, the resonance numbers change because the aggregation method changed, not because your voice did. See Goal profiles and the Methodology page.

Next: Interpreting your first result.