What is Vocal Analyzer?

Vocal Analyzer is a local, open-source tool that takes a short voice recording and produces a dashboard of acoustic measurements relevant to the perception of gender in voice. The analysis runs on your own device, and your recordings are not sent anywhere. It comes in two versions:

• The browser version runs entirely in your web browser. Nothing needs to be installed, and your recordings and results are stored in that browser.
• The installed app runs a local server on your computer, with Docker (make up) or the uv-managed dev server (make dev-serve), and adds a command-line tool and optional transcription and speaker similarity.

See Storage and privacy for where each version keeps your data.

Who it is for

• People exploring feminization, masculinization, or androgynization of their voice who want data to track alongside subjective listening.
• Voice-training hobbyists who want to understand why a recording sounds the way it does, not just whether it "passes".
• Researchers, curious engineers, and students who want a reproducible, inspectable stack that ties a dashboard value back to a literature- grounded measurement.

What it does

• Extracts pitch (fundamental frequency, F0), formants (F1/F2/F3), voice quality (spectral tilt, CPPS, LTAS complements), harmonicity (HNR), heuristic sibilant frication, intensity, and — in the installed app, when a reference is enrolled — speaker-embedding similarity.
• Overlays reference bands on the primary cues (pitch and a formant- dispersion resonance proxy) so a number like "175 Hz" lands in a visible context instead of floating alone.
• Flags measurement conditions that degrade a metric (low SNR, sample rate below 16 kHz, F0 above a per-metric reliability ceiling, formant-tracking instability) so the dashboard can soften its claims.
• Produces a timeline view for inspecting the recording frame by frame with pitch, formants, and — optionally, in the installed app — transcription overlays.
• Persists analyses locally so you can compare recordings over time.

What it does not do

• Not a clinical tool. The bands are population overlays. They are not validated predictors of listener judgement and must not be used as clinical cutoffs. Self-perception and listener perception are independent (Dacakis et al. 2017).
• Not a scoring machine. There is no single "feminine score" or "pass/fail" verdict. Every metric is reported descriptively, with explicit confidence framing when the recording conditions are unreliable.
• Not a replacement for a voice teacher. Acoustic measurements can not hear resonance shaping in real time, catch harmful compensations, or tell you what to practice next. A teacher can.
• Not a gender detector. The cues here correlate with perceived gender in English-speaking corpora. They do not determine anyone's gender, identity, or self-perception.

Language and scope limitations

• Measurement code and reference bands are calibrated against English-speaking adult corpora. Other languages — especially tonal languages — will behave differently.
• The per-vowel analysis uses an English vowel class set. Non-English vowels either group into the nearest class or are excluded.
• Intensity is uncalibrated: it is only comparable inside a single recording, not against other recordings or absolute dB SPL.

Next: Recording tips and Your first analysis.