Audio forensics
& voice comparison.

An original WAV or MP3 recording, or a video with an extractable audio track.

• Record confidence at the word level, rather than hiding uncertain words in fluent-looking text.
• Keep low-confidence words visibly bracketed.
• Return no words when speech is not discernible; silence or noise is not a failed transcription.
• Transcript release uses the platform’s verification state linked to the original recording.

Suggested text can influence what a listener subsequently hears. Treat the draft as a review aid, retain uncertainty and transcribe the original recording rather than an enhanced copy.

WAV, MP3 or M4A audio, or a video with an extractable audio track.

• Remove mains hum only at frequencies with a persistent narrowband signal.
• Reduce reverberation only when the recording is long enough to estimate a useful correction.
• Adapt noise-reduction strength to the recording’s signal quality.
• Level volume with noise-floor awareness, avoiding blind amplification of residual background noise.

Processing supports intelligibility and presentation, not proof of unheard content. Always transcribe the original, not this listening copy. The chain is non-generative and does not modify the sealed original.

A known recording and a questioned recording.

• Extract comparable voice characteristics from suitable speech in both recordings.
• Report the raw similarity between those characteristics.
• Screen both files for synthetic-voice risks before interpreting similarity.

No automatic same/different-speaker verdict. The supplied validation involved only three real speakers and is not a representative casework calibration. Cloned voices from a shared generation engine can appear unusually similar.

Two audio files or two video audio tracks.

• Compare a clip against its claimed longer source recording.
• Assess re-encoded versions or copies of the same underlying content.
• Compute and compare fingerprints locally without querying an external lookup service.

This asks whether the recorded content is shared, not whether the speaker is the same. Noise-dominated audio receives no match/no-match verdict because the signal is unreliable.

Audio or video audio with a detectable mains-hum trace above the required presence threshold.

• Check whether the hum signal exceeds a 10 dB peak-to-floor presence threshold.
• Run a coarse 16-second pass and a finer 2-second pass.
• Flag abrupt frame-to-frame changes for examination and report no trace when detection fails.

This is continuity checking only. There is no external grid-reference database, so it cannot establish an absolute recording date or time. A missing trace is not evidence of tampering; battery-powered or isolated recordings often lack one.

An audio recording or video audio track.

• Divide the recording into segments using detected speech.
• Compare adjacent segments for changes in speaker characteristics.
• Compare background-noise spectra to identify environment changes that voice comparison may miss.

A flag is not an automated splice verdict. Validation covers synthetic test data rather than a representative casework corpus. Genuine speaker changes or changing acoustic conditions require contextual review.

An audio file or a video carrying audio; confirm format support before submission.

• Extract container, codec and timestamp fields without a statistical model.
• Flag missing creation timestamps and available editing-software or encoder traces.
• Check for absent or multiple audio streams and document the available inventory.

Metadata is editable. A clean result does not establish that a recording is unaltered; a flagged result does not itself prove tampering. Combine file-record checks with content-level examination.