All comparisons

Deepgram vs OpenAI speech-to-text

Deepgram is ahead on time to first text and final-text delay, and OpenAI on the Pipecat Dataset. The Ocular Dataset is too close to call.

Each row compares the best Deepgram model with the best OpenAI model on that metric, from the same tests.

Last updated Methodology by Dileep Chagam

Best of each

Deepgram and OpenAI, metric by metric

MetricDeepgramOpenAI
Pipecat Dataset WER3.46%Deepgram Nova-32.16%GPT Realtime Whisper
Ocular Dataset WER4.18%Deepgram Flux English3.53%GPT Realtime Whisper
Time to first text791msDeepgram Flux English1.38sGPT Realtime Whisper
Final-text delay103msDeepgram Nova-3543msGPT Realtime Whisper
Deepgram line-up

Deepgram has 3 models on Converse-STT: Deepgram Flux English, Deepgram Flux Multilingual and Deepgram Nova-3. Its best on accuracy is Deepgram Nova-3 (3.46%). Its fastest is Deepgram Nova-3 (103ms).

OpenAI line-up

OpenAI has 3 models on Converse-STT: GPT-4o Mini Transcribe, GPT-4o Transcribe and GPT Realtime Whisper. Its best on accuracy is GPT Realtime Whisper (2.16%). Its fastest is GPT Realtime Whisper (543ms).

Verdict

Which to pick

Pick Deepgram for time to first text (588ms sooner) and final-text delay (440ms sooner). Pick OpenAI for the Pipecat Dataset (1.31 points lower). For the Ocular Dataset, the gap is inside the margin of error, so neither is ahead.

Frequently asked questions

Is Deepgram or OpenAI better for speech-to-text?

OpenAI is ahead, 1.31 points lower. The best Deepgram result is Deepgram Nova-3 at 3.46%, and the best OpenAI result is GPT Realtime Whisper at 2.16%, on the Pipecat Dataset.

Which is faster, Deepgram or OpenAI?

Deepgram is faster, 440ms sooner. The best Deepgram result is Deepgram Nova-3 at 103ms, and the best OpenAI result is GPT Realtime Whisper at 543ms, on final-text delay.

Which Deepgram and OpenAI models were tested?

Deepgram: Deepgram Flux English, Deepgram Flux Multilingual, Deepgram Nova-3. OpenAI: GPT-4o Mini Transcribe, GPT-4o Transcribe, GPT Realtime Whisper. Every model ran the same tests on Converse-STT.