Which speech recognition model handles noisy audio best?

On the leaderboard's noisy speech test (URGENT 2024), the lowest error rate belongs to zoom/scribe_v2_pro, a cloud API. OtterAI/speech-v3.1 is second, and the best open-weight model is CohereLabs/cohere-transcribe-03-2026. The average over the leaderboard's nine public test sets puts Otter first at 3.88 percent ¹ and Zoom at 4.08 percent ², so noise changes the order only at the very top. The reading is the version of October 17, 2026, taken on October 20, 2026.

Zoom and Otter lead on noisy speech, then AssemblyAI and ElevenLabs

In the Noisy speech (URGENT 2024) column of the tables below, Zoom scores lowest of all 56 data points ³. Otter is next, then AssemblyAI's universal-3-5-pro, then ElevenLabs Scribe, and then Cohere's open model.

Showing 8 of 13 rows. See every row in the dataset.

Every row, whatever its rank, scores worse on noisy speech than on its own average. Noise costs everyone accuracy. What changes between models is how much.

The best open model beats most cloud APIs on noise

Cohere's transcribe model has the lowest noisy-speech error rate of any open-weight model, and it beats nine of the 13 cloud APIs on that column, including Microsoft Azure AI Speech, OpenAI's gpt-transcribe and Gladia. Azure ranks above Cohere on the average (4.49 percent against 5.21 percent) ^{4 10}, yet behind it on noise. An average can hide this, which is why the column matters when your audio is not clean.

Showing 8 of 43 rows. See every row in the dataset.

Some models with a good average drop back on noise. nvidia/canary-qwen-2.5b has an average of 5.38 percent ¹⁶, but its noisy-speech error rate is well behind Cohere's. Read the column, not just the rank.

Our own babble-noise test agrees on direction

On the on-device noise test, we mixed background talk into 100 English recordings at set levels and ran three on-device engines on a Mac. Two of those engines are also on the leaderboard, so we can compare the ordering.

• Parakeet TDT v3 went from 4.44 percent ¹⁸ with no noise to 44.47 percent ¹⁹ with talk as loud as the speaker.
• Whisper large-v3-turbo went from 3.78 percent ²⁰ to 66.09 percent ²¹.
• Apple SpeechTranscriber went from 6.73 percent ²² to 71.88 percent ²³.

On the leaderboard, Parakeet v3 has a lower noisy-speech error rate than Whisper large-v3-turbo, and in our test it also holds up better at the loudest babble. The ordering on clean audio differs: in our test Whisper was ahead with no noise, while on the leaderboard's average Parakeet is ahead. So noise, not only the clean score, decides the winner. Our test also shows how steeply errors rise once the background is as loud as the speaker.

What the numbers do and do not measure

• The leaderboard's noisy set is degraded and noisy speech. It is not the same as a person talking in a busy cafe, and our babble is synthetic: it is built from read speech, which is steadier than a real room.
• Our test uses one language, one Mac and 15 data points ²⁴. Its finding is the ranking and the shape of the curve, not an exact error rate for your room.
• Leaderboard text is normalized before scoring, so capitals and punctuation are not counted. The audio is not dictation.
• Cloud APIs have no speed figure, and the leaderboard's own Average WER includes private recordings not published here.

What this means for dictating in a noisy place

Pick a model by its noisy-speech score if you dictate in an open office, a car or a cafe, and test it with your own microphone. Keeping your voice well above the room matters: our test shows how sharply errors rise when the room is as loud as you are. For a quiet room, the average is a fair guide, and the cloud leaders and top open models sit close together.

Related pages

• The full dataset and our noise test
• On-device speech recognition benchmark
• Open speech recognition models and speech-to-text APIs
• How benchmarks are built and the speech recognition overview