PERIT-ALIGN

The right word, a quarter-second late, is still the wrong caption.

PERIT-ALIGN scores the word-level timestamps a speech model returns — where each word starts and where it ends — against reference timings, so captions, redaction and audio search land on the right moment.

See the full leaderboardPut your model on the board
14
Models ranked
10
Providers compared
97.85%
Best timing accuracy @100 ms · Perit Golden Data Fine-tuned Model
19
Metrics reported per model
Results · 24 Sept 2026

Who won what.

Accuracy is only one of the questions, so each of these is the best model on its own axis.

Most accurate timing
97.85%Timing Accuracy @100 msPerit Golden Data Fine-tuned ModelPerit
Best speech-to-text timestamps
92.33%Timing Accuracy @100 mstranscribe-1Fish Audio
Tightest timing
83.90%Timing Accuracy @50 msPerit Golden Data Fine-tuned ModelPerit
Smallest boundary error
23.8 msBoundary MAEPerit Golden Data Fine-tuned ModelPerit
Best span overlap
84.98%Mean IoUtranscribe-1Fish Audio
Widest coverage
97.70%Word coveragemai-transcribe-2Microsoft
Timing accuracy @100 ms

Every model, ranked by how many words it timed right.

Higher is better. A word counts when both its start and its end land within 100 ms of the reference.

All 19 metrics
#Model
0%25%50%75%100%
@100 ms
  1. 01Perit Golden Data Fine-tuned ModelPerit97.85%
  2. 02transcribe-1Fish Audio92.33%
  3. 03voxtral-mini-transcribeMistral AI85.04%
  4. 04grok-stt-1.0xAI78.04%
  5. 05qwen3-asr-0.6bQwen77.41%
  6. 06qwen3-asr-1.7bQwen76.60%
  7. 07universal-3-5-proAssemblyAI63.45%
  8. 08nova-3Deepgram60.19%
  9. 09whisper-large-v3OpenAI59.89%
  10. 10mai-transcribe-2Microsoft59.85%
  11. 11whisper-large-v3-turboOpenAI46.95%
  12. 12whisper-1OpenAI45.86%
  13. 13parakeet-tdt-0.6b-v3NVIDIA30.96%
  14. 14nemotron-3.5-asr-streaming-multilingual-0.6bNVIDIA5.77%

Timing accuracy: the share of words placed within 100 ms of the reference. Longer bar is better. Perit's model is a forced aligner: it places the words of a known transcript in time, while the other models transcribe the audio and time the words in one pass.

How it was scored

Same audio, same reference timings, every model.

Each model's words are matched to the reference words and compared boundary by boundary — start against start, end against end. Coverage and the transcript's own WER are reported next to the timing, so a model cannot look precise by timing only the easy words.

Task
Word alignment · English
Models
14 from 10 providers
Ranking
Sorted by Timing Accuracy @100 ms descending, ties broken by boundary MAE.
Run date
24 Sept 2026
What the board reports
  • By tolerance. How many words and boundaries land inside each window, from 20 ms to 200 ms.
  • Boundary error. How far off the boundaries are, in milliseconds — the typical one, the worst tenth, starts against ends.
  • Drift & shape. Whether a model runs systematically early or late, and whether its word spans have real length.
  • Coverage & speed. How many words got a timestamp at all, how good the transcript underneath was, and request time — measured under concurrent load, so indicative only.
Method

How a run works.

  1. 01Every model gets the same audio. Speech-to-text models return word timestamps with their transcript; forced aligners place a known transcript in time.
  2. 02Returned words are matched to the reference words. Timing is scored on matched words, and coverage reports how many matched.
  3. 03Each boundary is compared on its own — start against start, end against end — and reported at 20, 50, 100 and 200 ms.
  4. 04Ranked by timing accuracy at 100 ms; ties are broken by mean boundary error.
  5. 05Latency is measured under concurrent load and is indicative only.