STT leaderboard: the speech-to-text SOTA.

The short answer: Granite Speech 4.1 2B is the current SOTA entry in this STT leaderboard, with 5.33% mean WER on the HF Open ASR Leaderboard. Pulse Pro leads among hosted APIs on accuracy, while Deepgram Nova-3 remains the go-to for low-latency streaming. For an open-source ASR model, start with Granite Speech 4.1 2B or Whisper Large v3 Turbo if multilingual coverage matters most.

LibriSpeech is still useful for clean-speech comparison, but it is saturated near the frontier. Use the eight-dataset Open ASR mean WER as the primary ranking, then check LibriSpeech clean/other when your workload is close-read English speech.

ASR models, ranked by mean WER.

Lower mean WER is better. Rows are ranked by mean WER across the eight HF Open ASR Leaderboard datasets, with each model's LibriSpeech test-clean WER shown alongside for a clean-speech reference point.

LibriSpeech ASR SOTA, in context.

LibriSpeech test-clean is a classic ASR benchmark, but it no longer separates the strongest speech-to-text systems by itself. Current frontier systems cluster around very low WER on clean audiobook speech, so CodeSOTA uses the broader HF Open ASR Leaderboard as the headline ranking and treats LibriSpeech as one diagnostic slice.

Use LibriSpeech when your workload is clean, read English audio. For meetings, calls, accents, long-form podcasts, or noisy streaming, check AMI, Earnings-22, GigaSpeech, TED-LIUM, VoxPopuli and latency features before choosing a model.

LibriSpeech test-clean WER from the HF Open ASR Leaderboard. Lower is better; the frontier is saturated near 1.3%, so treat sub-2% gaps as noise and rank on the eight-dataset mean WER above. Cloud-API rows show vendor- or AA-reported figures.

Open-source ASR or API?

The SOTA speech-to-text choice is not only the lowest WER. It is the trade between accuracy, latency, privacy, language coverage and operations.

Related speech pages.