CekuraBench

Compare speech-to-text models

Any two of the 17 models, head to head on accuracy and latency, measured under the same conditions.

Last updated Methodology by Dileep Chagam

Compare
AssemblyAI Universal 3.6 ProAssemblyAI
Soniox STT-RT v5Soniox
Pipecat Dataset WER
1.77%1st
0.78 points lower
2.54%8th
Ocular Dataset WER
4.50%14th
Too close to call
3.92%6th
Too close to call
Time to first text
485ms1st
710ms sooner
1.20s9th
Final-text delay
91ms3rd
38ms1st
53ms sooner
Lower is better on every metric. The small figure is the place in the whole field.All 136 comparisons
WER · Word error rate
Share of words the model got wrong: replaced, missing or added. Lower is better.
Pipecat Dataset
1,000 short public clips from Pipecat's STT benchmark dataset.
Ocular Dataset
8 recordings from 4 real phone conversations, licensed from Ocular.
TTFT · Time to first text
How long after the audio starts streaming until the first words appear.
TTFS · Final-text delay
How long after the caller stops until the full transcript arrives. A voice agent waits for this before it replies.

By provider

The best model from each provider, metric by metric.

Every comparison

AssemblyAI Universal 3.6 Pro

AssemblyAI Universal 3.5 Pro

Google Chirp 3

Reson8

GPT Realtime Whisper

Cartesia Ink 2

Speechmatics Linden

Soniox STT-RT v5

Google Chirp 2

Inworld STT-1

GPT-4o Mini Transcribe

Deepgram Nova-3

Smallest Pulse

GPT-4o Transcribe

Deepgram Flux English

Deepgram Flux Multilingual

Gradium

Questions

Speech-to-text model comparison FAQ

How do I compare speech-to-text models and APIs?

Choose any two models to compare word error rate, time to first text, and final-text delay from the same recordings and test pipeline. Lower values are better for every displayed metric.

Which speech-to-text model is the most accurate?

Accuracy depends on the audio and dataset. Use the Pipecat and Ocular word error rates separately, because a model can rank differently on clean clips and real conversations.

Can I compare Deepgram with AssemblyAI?

Yes. Select the tested Deepgram and AssemblyAI models to compare accuracy and streaming latency under the same conditions, including direct model-version comparisons.

What matters besides word error rate?

Voice agents also depend on how quickly partial text arrives, how soon the final transcript is available after the speaker stops, whether text streams during speech, and whether the provider returns usable output consistently.