BAIAN / INTERNAL LISTENING REVIEW

Listen. Check the speaker changes.

Ten short English and Arabic excerpts from the Nemotron evaluation. Listen first, then open the predictions and note what it got right or wrong.

These are original recordings with predicted speaker labels, not synthetic voice auditions. Human accuracy review is still pending.

Notes stay in this browser. Export them before switching devices.

Read the GPU test report

CLIP 01 · 10 SECONDS

Opening change; English control

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–0.98s · Speaker 1
  • 1.30–3.10s · Speaker 1
  • 3.61–10.00s · Speaker 2

GPU predictions · 25 September

Separate L4 run on this 10-second clip. Times start at zero in this player. Speaker numbers are local to each run and do not establish matching identities across runs.

  • 0.00–3.19s · Speaker 1
  • 3.62–9.99s · Speaker 2

Processing: 1.63s. Model loading: 24.59s. This measures processing speed, not live delay.

Original source · source interval 330–340s · GPU run

Both sets are unscored predictions. Different amounts of preceding audio mean this is not a controlled CPU-versus-GPU accuracy comparison.

CLIP 02 · 17 SECONDS

Returning first speaker; English control

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–4.24s · Speaker 2
  • 4.78–12.37s · Speaker 1
  • 13.12–13.72s · Speaker 1
  • 14.74–16.99s · Speaker 1

CLIP 03 · 18 SECONDS

Brief second-label/overlap predictions; verify real interjection vs false speaker

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–5.52s · Speaker 1
  • 4.98–5.80s · Speaker 2
  • 5.80–18.00s · Speaker 1
  • 13.07–13.08s · Speaker 2
  • 13.09–13.24s · Speaker 2

GPU predictions · 25 September

Separate L4 run on this 18-second clip. Times start at zero in this player. Speaker numbers are local to each run and do not establish matching identities across runs.

Please judge the brief second-speaker detections at 5.05–5.81s and 13.05–13.27s. Are these real interruptions or false detections?

  • 0.00–4.31s · Speaker 1
  • 4.32–5.41s · Speaker 1
  • 5.05–5.81s · Speaker 2
  • 6.10–17.99s · Speaker 1
  • 13.05–13.27s · Speaker 2

Processing: 2.03s. Model loading: 16.55s. This measures processing speed, not live delay.

Original source · source interval 1829–1847s · GPU run

Both sets are unscored predictions. Different amounts of preceding audio mean this is not a controlled CPU-versus-GPU accuracy comparison.

CLIP 04 · 19 SECONDS

Short second label then continued main voice

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–6.82s · Speaker 1
  • 5.19–5.59s · Speaker 2
  • 7.75–8.59s · Speaker 1
  • 9.44–16.06s · Speaker 1
  • 16.73–19.00s · Speaker 1

CLIP 05 · 20 SECONDS

Potential interruptions and overlap

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–5.35s · Speaker 1
  • 5.35–6.07s · Speaker 2
  • 5.99–19.74s · Speaker 1
  • 11.24–11.55s · Speaker 2

CLIP 06 · 20 SECONDS

Ordinary continuous speech comparison

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–4.19s · Speaker 1
  • 5.03–20.00s · Speaker 1

CLIP 07 · 20 SECONDS

Ordinary Arabic speech; check missed speech

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–1.84s · Speaker 1
  • 2.97–5.87s · Speaker 1
  • 5.96–9.05s · Speaker 1
  • 9.50–11.79s · Speaker 1
  • 12.93–15.25s · Speaker 1
  • 15.82–20.00s · Speaker 1

CLIP 08 · 20 SECONDS

Check within-excerpt identity stability

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–1.84s · Speaker 1
  • 3.39–13.03s · Speaker 1
  • 13.05–13.06s · Speaker 1
  • 13.07–13.11s · Speaker 1
  • 13.12–13.18s · Speaker 1
  • 13.20–13.27s · Speaker 1
  • 13.28–20.00s · Speaker 1

CLIP 09 · 35 SECONDS

Opening multiple labels; inspect actual voices and edited footage

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 4.03–10.17s · Speaker 2
  • 10.26–22.06s · Speaker 3
  • 22.00–31.20s · Speaker 4

CLIP 10 · 26 SECONDS

Opening label transition from speaker4 to5; verify identity

CPU predictions · 24 September

Earlier CPU run on a longer excerpt; times clipped and rebased to this player. Labels are unverified.

  • 0.00–9.16s · Speaker 5
  • 9.20–10.49s · Speaker 5
  • 11.17–23.40s · Speaker 6
  • 23.60–23.66s · Speaker 6
  • 23.99–26.00s · Speaker 6