Hear exactly where a delivery goes wrong.

SpeechLens lines a participant's reading up against a reference reading of the same text, word by word, and marks each delivery flaw with its time span, the numbers behind it, and a fix.

Pick a dataset clip on the left to see it work.

The contrastive dataset

Each reference speech is mirrored by recordings of the same transcript with delivery flaws, from near-perfect (severity 1) to egregious (severity 5). Synthetic flaws are injected with PSOLA resynthesis in the speaker's own voice, so every label is exact to the sample. Human recordings were made by the team following a scripted flaw sheet.

Speech Kind Flaw

Clip	Kind	Flaws	Length

How SpeechLens works

Judging a delivery means comparing it with how the same words should sound. SpeechLens makes that comparison explicit: both readings are aligned word by word, measured on the same acoustic signals, and every difference beyond the reference speaker's own natural variation becomes a finding with a time span, the numbers behind it and a fix.

1 · Align

An exact CTC Viterbi pass over wav2vec2 character probabilities finds where every word of the transcript starts and ends, in both readings. Boundaries are refined on the energy envelope.

2 · Measure

Pitch in semitones relative to each speaker's own median, loudness relative to their own level, energy above 2 and 4 kHz, pauses, held voicing, recogniser confidence. A deep and a high voice reading identically look identical.

3 · Contrast

Each signal is split into the speaker's overall habit and the local deviation, scored against the reference speaker's word-to-word variation. Data-fitted gates drop anything within natural reread variation.

4 · Explain

Every finding names one measured cause, how far it moved and why it matters to a listener, with a fix. No language model: the same audio always gives the same words.

5 · Prove it

Hear it fixed moves only that cause back to the reference value in the participant's own voice. Re-analysed, the finding disappears in 91 % of cases: the explanation is causal, not a guess.

6 · Practise

Record a retake of one passage and see it compared with the original against the same reference. Track progress across takes.

What makes it different

Contrastive by design. Remove the reference and the same detector falls from F1 0.68 to 0.07; normal expressive variation looks like a flaw without it.
Exact ground truth. A flaw injection engine edits real speeches with PSOLA in the speaker's own voice, so every label is exact to the sample: 350 synthetic clips, 11 flaw types, 5 severities.
Tested on new voices, people and stress. F1 0.85 on 20 unseen speakers (10 women) with frozen thresholds; team recordings with a manipulation check and a blind listening audit; MP3, telephone, echo, tempo and noise conditions; byte-identical reruns.
Honest about limits. Background noise the reference lacks breaks the comparison; the app measures it and warns.