Open ASR Leaderboard: Speech Recognition Accuracy, Tracked

Paid cloud APIs make the fewest errors on the Open ASR Leaderboard's public English tests, but the best open models trail them by a small margin. OtterAI/speech-v3.1 averages 3.88 percent word error rate ¹, followed by zoom/scribe_v2_pro at 4.08 percent ². The strongest open-weight model, Qwen/Qwen3-ASR-1.7B-hf, reaches 4.94 percent ³. Below that, the field bunches tightly, and the real separation comes from the type of audio, not from the label "open" or "paid".

The table holds 56 data points ⁴. Each is one model's average word error rate across the leaderboard's public English test sets, with its score on every individual set. Lower is better. For background on how the metric works, see our guide to speech recognition benchmarks and the overview of speech recognition.

The top of the table is cloud APIs, led by a narrow margin

The cloud APIs fill the first places. After Otter and Zoom come modulate/multilingual at 4.53 percent ⁵, microsoft/azure-speech-07-2026 at 4.49 percent ⁶ and elevenlabs/scribe_v2 at 4.55 percent ⁷. The differences between these neighbors are fractions of a point. A different set of test recordings could reorder them.

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

Not every paid API is at the top. openai/gpt-transcribe averages 5.49 percent ¹¹, and gladia/solaria-3 averages 5.59 percent ¹². Both sit in the same band as the strong open models. Paying for an API does not guarantee the lowest error rate on these tests. See our overview of speech-to-text APIs for the market side.

The best open models are within reach of the paid ones

The open-weight rows are led by Qwen3-ASR-1.7B at 4.94 percent ³. A cluster follows closely: CohereLabs/cohere-transcribe-03-2026 at 5.21 percent ¹³, bosonai/higgs-audio-v3-stt at 5.22 percent ¹⁴ and TheStageAI/thewhisper-large-v3-turbo at 5.24 percent ¹⁵. Several of these would outscore some well-known paid APIs.

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

Open weights also come with a practical advantage the error rate does not capture: you can run them on your own machine. Our overview of open speech recognition models covers that segment, and the on-device ASR benchmark shows what happens when such models run on consumer hardware instead of the leaderboard's own.

Size does not buy accuracy, and speed is a separate choice

Bigger models are not better on this table. mistralai/Voxtral-Small-24B-2507 is by far the largest open model on the table, yet it averages 5.81 percent ²⁰. nvidia/parakeet-tdt-0.6b-v2 is a small fraction of that size and averages 5.45 percent ²¹.

Speed varies far more than accuracy among the good models. The leaderboard reports a speed figure (RTFx, seconds of audio processed per second of compute) for every open model, shown in the speed column of the full table. ibm-granite/granite-speech-5.0-470m-turboctc scores 5.78 percent ²² and sits among the fastest rows, while Voxtral-Small-24B is among the slowest. If you need transcription of large archives, speed matters as much as the last tenth of a point.

The leaderboard also keeps separate rows for one model run with a faster setup. nvidia/parakeet-tdt-0.6b-v2 (fast-gpu-asr) scores 5.44 percent ²³, essentially the same as the standard row, but its speed figure is much higher. Decoding choices can change accuracy more. The soundsgoodai Zipformer model scores 6.22 percent with transducer search ²⁴ and 7.64 percent with greedy CTC search ²⁵.

Tiny models pay a clear price. The abr-ai Niagara family spans from very small models upward, and its smallest row, abr-ai/niagara-9m-batch.en, averages 15.27 percent ²⁶, the worst on the table. Its larger sibling niagara-84m-batch.en improves to 8.87 percent ²⁷. These models trade accuracy for a very small footprint.

Meetings and noise separate the models far more than audiobooks do

The average hides large differences between recordings. On read audiobooks (LibriSpeech clean), nearly every serious model is near the floor: the leading cloud and open models all score close to zero errors per hundred words. The table lists every score per set. The ordering among the leaders on that set tells you little about real use.

Meetings and noisy audio are where models diverge. On AMI meetings, zoom/scribe_v2_pro is lowest of the models highlighted here, while openai/whisper-large-v3-turbo and facebook/omniASR-LLM-7B-v2 are more than twice as bad on the same recordings. Whisper large-v3 averages 6.66 percent ²⁸ and its turbo variant 7.17 percent ²⁹. The faster variant's weakness is concentrated in a few sets, notably European Parliament speech.

On noisy speech (URGENT 2024), Zoom and Otter again lead, and the best open models, such as Cohere's, are a visible step behind. Noise is where paid APIs keep their clearest advantage. For a closer look, see our noise robustness data and the wider accuracy comparison.

Where the sets disagree

No single model wins everywhere. assemblyai/universal-3-5-pro averages 4.87 percent ¹⁰ overall, but its meeting score is much weaker than Zoom's. elevenlabs/scribe_v2 is among the best on parliamentary speech yet falls back on meetings. Cohere's open model is excellent on audiobooks and noise but loses ground on European Parliament speech, where several Whisper-family and Parakeet rows do better.

The consequence: a two-place gap in the average between neighbors says little, and a model that is second on the average may be first on the kind of audio you record. Look at the column that matches your use.

What this table cannot tell you

It is not dictation. The test sets are read audiobooks, meetings, earnings calls, podcasts, parliament speeches, noisy speech and Indian English conversations. A person dictating an email to a microphone in a quiet room is closest to the audiobook sets, where everyone scores well.

The average differs slightly from the leaderboard's own. The leaderboard's Average WER also includes two averages over private recordings that are not published, so its ranking can differ slightly from the public-set average used here.

Punctuation and capitals do not count. Text is normalized before scoring. A model that writes messy punctuation can score the same as one that writes clean text, which matters if you want to send the output without editing.

Speed is measured on the leaderboard's hardware. Cloud APIs have no speed figure, and the open models' speed will differ on your laptop or phone. Latency, price, privacy and language coverage are not measured at all.

Rankings move weekly. This is one reading of a table that is updated often, and new entries, such as the preview row sophea/asr-k1 at 4.61 percent ⁸, can change the order. Several rows are previews or lightly documented entries, and the leaderboard cannot verify how a cloud API changes behind its endpoint.

What this means if you choose how to write

For the lowest error rate on messy audio, use a top cloud API. Meetings, calls and noisy recordings are where the leaders open a real gap. Check the per-set columns for the audio you record.

For private or offline use, an open model is now a sound choice. Qwen3-ASR-1.7B, Cohere's transcribe model and the Parakeet family give up little on average, and they run on your own hardware. Parakeet in particular pairs strong scores with high speed.

For everyday dictation, small differences in this table will not decide your experience. Punctuation, correction tools, latency and how the app handles your accent matter more than a point of average error. If your voice is typically recorded cleanly, nearly every model in the upper half of the table will give you usable text.

Match the test to your voice. If you speak Indian English, compare the Indian English column, where the leaders are well below their own averages. If you work in noisy places, weigh the noise column most. Our other pages on noise robustness and on-device accuracy give more detail for those cases.