ConductSpeech

Benchmarks

How close we get to the right answer

We ran ConductSpeech on 290 public recordings and compared its transcripts with the words each person read, and its pitch with Praat's.

  • Praat
  • ConductSpeech since version 2.31.0
  • ConductSpeech before version 2.31.0

Pitch across 6 seconds of one recording. The gray dots are pitch the old version reported where Praat detects no voicing.

Pitch

Pitch is now within 8% of Praat

Praat is the free program voice researchers use to measure pitch. Before version 2.31.0, ConductSpeech also measured pitch in moments Praat doesn't count as voice, so pitch variation and range came out 34% to 55% higher than Praat's. We fixed that and checked the fix on 205 recordings we hadn't used to build it.

  • Before version 2.31.0
  • Since version 2.31.0
  • Praat
Pitch variation in semitones
Pitch range in semitones
All pitch numbers
MeasureBefore 2.31.0Since 2.31.0Praat
Children, 91 recordings
Mean pitch (Hz)232.6226.8224.5
Variation (semitones)5.023.483.36
Range (semitones)14.318.799.21
Second-language women, 72 recordings
Mean pitch (Hz)230.1226.5223.2
Variation (semitones)4.913.683.67
Range (semitones)14.089.479.71
Second-language men, 42 recordings
Mean pitch (Hz)156.3147.1143.2
Variation (semitones)5.684.213.91
Range (semitones)15.3010.649.86

Median pitch since 2.31.0, gap to Praat: Children 0.14%, Second-language women 0.3%, Second-language men 0.57%.

Samples analyzed before Nov 6, 2026 may still show the older, higher pitch numbers. Upload the recording again to update them.

205 recordings in 3 groups · Nov 5, 2026 · version 2.31.0 · group means · one person reading aloud in each recording

Transcripts

Transcripts have more errors for second-language speakers

We compared each transcript with the text the speaker read and counted the missed, added and wrong words. The second-language speakers are children and adults whose first language is Mandarin.

Word error rate by group

Native adults

4.9%

Median recording 4.1%

Children

22.2%

Median recording 19.7%

Second-language women

26.0%

Median recording 14.6%

Second-language men

26.1%

Median recording 13.2%

Each dot is one recording, and the dark line is the median. The right edge holds 2 recordings with more errors than words read.

Listen back before you score

The editor plays the recording one line at a time and underlines the words the transcriber wasn't sure of.

See the transcript editor

290 recordings · Nov 6, 2026 · version 2.31.1 · one person reading aloud in each; conversation is harder to transcribe

Voice measures

Which voice measures you can compare with Bridge2AI's

Bridge2AI-Voice publishes the code behind its voice measures (senselab, run with b2aiprep's settings). We ran that code and ConductSpeech on the same 290 recordings and compared each measure.

Same scale2 of 12

Compare the values directly.

  • Recording length
  • Mean loudness
Needs an offset1 of 12

Correct for the steady difference, then compare.

  • Mean pitch
Direction only3 of 12

Use for rankings and group differences only.

  • Harmonics to noise (HNR)
  • Cepstral peak prominence (CPPS)
  • Articulation rate
Not comparable5 of 12

Compare only with other ConductSpeech results.

  • Signal to noise
  • Voiced share
  • Loudness variation
  • Pause rate
  • Mean pause
Not tested here1 of 12

None of these recordings clipped, so it couldn't be checked.

  • Clipping

290 public recordings, none from Bridge2AI · Nov 5, 2026 · version 2.31.0

Limits

What we haven't tested yet

Children speaking their first language

Every child here was a Mandarin speaker reading English.

Telling the child's voice from the adult's

Speaker labels from voices stay off until we check them against clinicians' corrections.

Language measures against hand scoring

MLU and the other language measures are checked against software counts on 20 samples, and against counts that follow SALT's written conventions on made-up samples. Checks against clinicians' hand scoring of real samples come next.

How we ran the tests

The recordings

290 public recordings of people reading aloud, free to use under CC BY 4.0. The children and adults whose first language is Mandarin come from SpeechOcean762 (Zhang and colleagues, 2021; OpenSLR 101). The adults reading English audiobooks come from LibriSpeech dev-clean (Panayotov and colleagues, 2015; OpenSLR 12).

The rules

  • We set the scoring rules before each run.
  • We picked the recording at the top by a fixed rule: of the 72 second-language women, the one with the median old error in pitch variation.
  • Transcripts came from the settings every upload uses.
How we sorted the voice measures

By the correlation (r) between our numbers and Bridge2AI's, in the group where they agree least:

  • Same scale: r of 0.90 or more, and in every group an average difference no more than a quarter of Bridge2AI's standard deviation.
  • Needs an offset: r of 0.90 or more, with a larger average difference.
  • Direction only: r from 0.60 to 0.90.
  • Not comparable: r below 0.60.

Versions

TestResultsDateVersion
Pitch1.1.0Nov 5, 20262.31.0
Voice measures1.1.0Nov 5, 20262.31.0
Transcripts1.0.0Nov 6, 20262.31.1

Page data 1.2.0.

ConductScience ran these tests. ConductSpeech has not been independently validated in a peer-reviewed study.

Try it on your own recordings

Check the transcript against what you hear, then read the measures.