Preprint

Robust Speech Recognition via Large-Scale Weak Supervision

Alec Radford et al., arXiv (eess.AS), 2022

Read the original

Read on October 16, 2026

What it is: Introduces Whisper, a speech recognition model trained on 680,000 hours of multilingual audio from the internet. The authors report zero-shot results on standard benchmarks and compare the models with human transcribers.
How much weight it carries: A preprint by the OpenAI team, widely used and reproduced. It reports benchmark results, not real-world dictation in noisy rooms, and the comparison with humans depends on the benchmark chosen.

Facts taken from this source

OpenAI's Whisper speech recognition models were trained on 680,000 hours of multilingual and multitask audio with transcripts.
“When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks” (Abstract)
On the AMI meeting corpus recorded with individual headset microphones (IHM), Whisper Large V2 had a word error rate of 16.9 percent.
“CORAAL 35.6 16.2 54.5 AMI IHM 37.0 16.9 54.3 Switchboard 28.3 13.8 51.2” (Table 2, effective robustness across datasets, AMI IHM row)
On the AMI meeting corpus recorded with a single distant microphone (SDM1), Whisper Large V2 had a word error rate of 36.4 percent.
“AMI SDM1 67.6 36.4 46.2 LibriSpeech Other 6.2 5.2 16.1” (Table 2, effective robustness across datasets, AMI SDM1 row)
Radford and colleagues report that, compared with humans, their Whisper speech recognition models approach human accuracy and robustness.
“When compared to humans, the models approach their accuracy and robustness.” (Abstract)
On the CHiME6 test set, Whisper Large V2 had a word error rate of 25.5 percent.
“CHiME6 65.8 25.5 61.2 VoxPopuli En 17.9 7.3 59.2” (Table 2, effective robustness across datasets)
On the LibriSpeech Clean test set, Whisper Large V2 had a word error rate of 2.7 percent.
“LibriSpeech Clean 2.7 2.7 0.0 Artie 24.5 6.2 74.7” (Table 2, effective robustness across datasets)
The 2022 paper that introduced Whisper lists six authors, among them Greg Brockman and Ilya Sutskever.
“Alec Radford , Jong Wook Kim , Tao Xu , Greg Brockman , Christine McLeavey , Ilya Sutskever” (Author list on the arXiv abstract page)
On the Switchboard test set, Whisper Large V2 had a word error rate of 13.8 percent.
“Switchboard 28.3 13.8 51.2 CallHome 34.8 17.6 49.4” (Table 2, effective robustness across datasets)
The Whisper paper says its evaluation audio covers recording conditions including telephone and VoIP calls and meetings.
“The audio selection covers various recording conditions such as scripted and unscripted broadcast, telephone and VoIP calls, and meetings.” (Evaluation data description)

Robust Speech Recognition via Large-Scale Weak Supervision

Facts taken from this source

Pages that rest on this source

Robust Speech Recognition via Large-Scale Weak Supervision

Facts taken from this source

Pages that rest on this source