Ten minutes of transcribed speech was enough for wav2vec 2.0 to reach a 4.8% word error rate on the LibriSpeech test-clean set (Meta AI, wav2vec 2.0, NeurIPS 2020), after the model first trained on raw audio with no human transcripts, the approach known as self-supervised learning. On the hidden test set of the SUPERB @ SLT 2022 challenge, WavLM-large scored 1242 on a scale where the October 2021 state of the art sits at 1000 (Feng et al., SUPERB @ SLT 2022), yet only about half of 33 benchmarked models beat a classic baseline on every task (Yang et al., IEEE/ACM TASLP 2024). The older checkpoints remain in daily use: facebook/wav2vec2-base-960h showed 1,737,874 downloads last month (Hugging Face, October 31, 2026). We aggregated data from the SUPERB, SUPERB-SG and ML-SUPERB benchmark papers, the SUPERB @ SLT 2022 challenge report, the wav2vec 2.0, WavLM and XEUS model papers, and live Hugging Face download counters. What these low error rates leave out is covered in our roundup on speech recognition hallucination rates.
A note on reading every number below: SUPERB scores are mostly from 2021 to 2024 because the benchmark papers set the record, and the live leaderboard could not be reached for this update (see Data watch). Lower is better for error rates; higher is better for accuracy and for SUPERB scores.
TL;DR
- wav2vec 2.0 Large hit 4.8% / 8.2% WER with 10 minutes of labels, against 16.3% / 25.2% for the previous best method (Meta AI, wav2vec 2.0, NeurIPS 2020).
- wav2vec-U reached 5.9% WER on test-other with zero transcribed speech (Meta AI, wav2vec-U 2021).
- The IEEE review’s state-of-the-art table lists 1.4% / 2.6% LibriSpeech WER with self-supervised pretraining, against 2.1% / 4.0% without (Mohamed et al., IEEE JSTSP 2022).
- HuBERT Large cut phoneme error rate to 3.53%, from 82.01% for the FBANK baseline (Yang et al., SUPERB, Interspeech 2021).
- Emotion recognition remains the low ceiling, with XEUS on top at 71.08% accuracy (Chen et al., XEUS, EMNLP 2024).
- Speech translation BLEU rose from 2.32 with FBANK to 20.01 with HuBERT Large (Tsai et al., SUPERB-SG, ACL 2022).
- About half of 33 models beat FBANK on all 15 tasks (Yang et al., IEEE/ACM TASLP 2024).
- WavLM Base+ scored 73.4 overall at 94.70M parameters, above HuBERT Large at 72.2 with 316.61M (Microsoft, WavLM, IEEE JSTSP 2022).
- The largest challenge model, at 619.82M parameters, scored 848 against 1242 for WavLM-large at 315.45M (Feng et al., SUPERB @ SLT 2022).
- On the 10-minute ML-SUPERB set, wav2vec2-large scored 598.3 and the smaller wav2vec2-base 755.2 (Shi et al., ML-SUPERB, Interspeech 2023).
- DistilHuBERT makes HuBERT 75% smaller and 73% faster (Chang et al., ICASSP 2022).
- microsoft/wavlm-large showed 1,068,711 downloads last month (Hugging Face, October 31, 2026).
1. Labeled Data: From 960 Hours to 10 Minutes
Word error rate (WER) is the share of words a system gets wrong, and “labels” here means human transcripts. LibriSpeech has two standard test sets: test-clean and the harder test-other. Going from 960 hours of labels to 10 minutes raised wav2vec 2.0’s test-other WER from 3.3% to 8.2%. Since 960 hours is 57,600 minutes, that is 5,760 times less transcribed audio (57,600 / 10) for an error rate about 2.5 times higher (8.2 / 3.3).
The bottleneck moved. Transcription used to set the floor on quality; after wav2vec 2.0, the scarce input became unlabeled audio and GPU time. Its 8.2% on test-other also sits about 67% below the 25.2% of the prior best method ((25.2 - 8.2) / 25.2), and the IEEE review credits it as the first model to reach single-digit WER on LibriSpeech using only the low-resource Libri-light subsets.
| Metric | Value | Source |
|---|---|---|
| WER with 10 minutes of labels, wav2vec 2.0 Large vs previous best Discrete BERT (test-clean / test-other) | 4.8% / 8.2% vs 16.3% / 25.2% | Meta AI, wav2vec 2.0, NeurIPS 2020 |
| WER with 10 minutes of labels, HuBERT X-Large (test-clean / test-other) | 4.6% / 6.8% | Meta AI, HuBERT, IEEE/ACM TASLP 2021 |
| WER on test-other with 1 hour of labels, wav2vec 2.0 Large vs WavLM Large | 5.8% vs 5.1% | Microsoft, WavLM, IEEE JSTSP 2022 |
| WER with 10 hours of labels, HuBERT Large and WavLM Large (both test-clean / test-other) | 2.4% / 4.6% | Microsoft, WavLM, IEEE JSTSP 2022 |
| WER with all 960 hours of LibriSpeech labels, wav2vec 2.0 Large (test-clean / test-other) | 1.8% / 3.3% | Meta AI, wav2vec 2.0, NeurIPS 2020 |
| WER on test-other with zero transcribed speech, wav2vec-U | 5.9% | Meta AI, wav2vec-U 2021 |
| LibriSpeech WER in the review’s state-of-the-art table, without vs with self-supervised pretraining (test-clean / test-other) | 2.1% / 4.0% vs 1.4% / 2.6% | Mohamed et al., IEEE JSTSP 2022 |
The comparison is not perfectly like for like. Discrete BERT used a 4-gram language model, while the wav2vec 2.0, HuBERT and WavLM rows above use a Transformer language model at decoding. The self-supervised entry in the last row combines a conformer XXL model with noisy student self-training. Most recent available data: Meta AI and Microsoft papers, 2020 to 2022.
2. SUPERB Content Tasks: Recognizing What Was Said
SUPERB started with 10 tasks and one strict rule: the pretrained model stays frozen and only a small task head trains. The baseline is FBANK, classic filter-bank audio features with no pretraining at all. HuBERT Large took phoneme error rate from 82.01% to 3.53%. A phoneme error rate is the same idea as WER, counted on speech sounds instead of words.
The headline gaps overstate one thing. The paper notes FBANK cannot work on the tasks limited to a linear head, which explains scores as low as 8.63% for keyword spotting. With a non-linear head on ASR, FBANK is competitive enough that its 23.18% is a fairer floor. Query by example, which finds a spoken term from an audio sample and is scored by maximum term weighted value (MTWV), posts the lowest absolute numbers of the group: the top score in the original paper was 0.0736.
| Metric | Value | Source |
|---|---|---|
| Phoneme error rate on LibriSpeech, FBANK baseline vs HuBERT Large | 82.01% vs 3.53% | Yang et al., SUPERB, Interspeech 2021 |
| Keyword spotting accuracy on Speech Commands, FBANK vs wav2vec 2.0 Large | 8.63% vs 96.66% | Yang et al., SUPERB, Interspeech 2021 |
| Automatic speech recognition WER without a language model, FBANK vs HuBERT Large | 23.18% vs 3.62% | Yang et al., SUPERB, Interspeech 2021 |
| Query by example MTWV score, FBANK vs HuBERT Base | 0.0058 vs 0.0736 | Yang et al., SUPERB, Interspeech 2021 |
| Automatic speech recognition WER on SUPERB, WavLM Large vs XEUS | 3.44% vs 3.34% | Chen et al., XEUS, EMNLP 2024 |
| Keyword spotting accuracy on SUPERB, WavLM Large vs XEUS | 97.86% vs 98.32% | Chen et al., XEUS, EMNLP 2024 |
The most recent English SUPERB numbers here come from XEUS, and the margins have become thin: 0.10 percentage points on ASR. The task family looks close to saturated.
3. Speaker, Intent, and Emotion Tasks
Speaker tasks ask who is talking rather than what was said, so they test whether a representation kept more than the words. Speaker identification climbed from effectively zero with FBANK to 90.33% with HuBERT Large. Fine-tuned WavLM Large reached a 0.383% equal error rate on VoxCeleb1-O, against 1.010% for the ECAPA-TDNN reference system. Equal error rate (EER) is the operating point where false accepts equal false rejects, so lower is better.
Emotion is the outlier. Four-class emotion accuracy on IEMOCAP has barely moved between WavLM Large and XEUS, and even the best score leaves almost 3 in 10 clips wrong. The practical limits are covered in our speech emotion recognition accuracy statistics. The diarization rows use different datasets, so see our speaker diarization error rate roundup before comparing them. Diarization error rate (DER) measures how much audio time gets “who spoke when” wrong.
| Metric | Value | Source |
|---|---|---|
| Speaker identification accuracy on VoxCeleb1, FBANK vs HuBERT Large | 8.5E-4 vs 90.33 (percent) | Yang et al., SUPERB, Interspeech 2021 |
| Intent classification accuracy on Fluent Speech Commands, FBANK vs HuBERT Large | 9.10% vs 98.76% | Yang et al., SUPERB, Interspeech 2021 |
| Intent classification accuracy in the review’s state-of-the-art table, without vs with self-supervised pretraining | 98.8% vs 99.3% | Mohamed et al., IEEE JSTSP 2022 |
| Speaker diarization DER on two-speaker LibriMix, WavLM Large vs XEUS | 3.24% vs 3.11% | Chen et al., XEUS, EMNLP 2024 |
| Emotion recognition accuracy on IEMOCAP (four classes), WavLM Large vs XEUS | 70.62% vs 71.08% | Chen et al., XEUS, EMNLP 2024 |
| Speaker verification EER on VoxCeleb1-O, ECAPA-TDNN reference vs WavLM Large with large-margin fine-tuning and score calibration | 1.010% vs 0.383% | Microsoft, WavLM, IEEE JSTSP 2022 |
| Diarization DER on CALLHOME telephone calls, EEND-EDA clustering vs WavLM Large | 11.84% vs 10.35% | Microsoft, WavLM, IEEE JSTSP 2022 |
Note the change of method in the speaker verification row: that result comes from fine-tuning, not from SUPERB’s frozen setup, so it is not a leaderboard score. The IEEE review also shows how small the edge can get at the top: 0.5 percentage points on intent classification.
4. SUPERB-SG and the 15-Task Benchmark: Where Self-Supervision Stops Helping
SUPERB-SG added five harder tasks in semantics and generation. Translation shows the biggest jump: BLEU, a translation quality score, rose from 2.32 to 20.01. Voice conversion also improves sharply, measured here by how many words a recognizer gets wrong on the converted audio.
Then the gains stop. Enhancement and separation depend on low-level signal detail, and the authors found those two tasks are not strongly correlated with the rest. wav2vec 2.0 Large actually sits below FBANK on the PESQ speech quality score. The journal version confirms it at scale: a model built to summarize speech is not automatically good at reconstructing it. For the separation side of this picture, see our speech separation statistics.
| Metric | Value | Source |
|---|---|---|
| Speech translation BLEU on CoVoST2 (English to German), FBANK vs HuBERT Large | 2.32 vs 20.01 | Tsai et al., SUPERB-SG, ACL 2022 |
| Out-of-domain speech recognition WER (Spanish, Mandarin, Arabic and spontaneous English), FBANK vs HuBERT Large | 63.58% vs 44.08% | Tsai et al., SUPERB-SG, ACL 2022 |
| WER of voice-converted speech on VCC2020, FBANK vs HuBERT Large | 38.3% vs 9.0% | Tsai et al., SUPERB-SG, ACL 2022 |
| Speech separation SI-SDRi score (signal quality improvement, higher is better), FBANK vs HuBERT Large | 9.23 vs 10.45 | Tsai et al., SUPERB-SG, ACL 2022 |
| Speech enhancement PESQ score on Voicebank-DEMAND, FBANK vs wav2vec 2.0 Large vs HuBERT Large | 2.55 vs 2.52 vs 2.64 | Tsai et al., SUPERB-SG, ACL 2022 |
| Correlation of task results with ASR, for speech translation and for out-of-domain ASR | 0.92 each | Tsai et al., SUPERB-SG, ACL 2022 |
| Share of the 33 benchmarked models that beat FBANK on all 15 tasks | About half | Yang et al., IEEE/ACM TASLP 2024 |
| Tasks where the remaining models trail FBANK | Around 1 to 3 of 15, most often enhancement and separation | Yang et al., IEEE/ACM TASLP 2024 |
The out-of-domain row deserves attention: HuBERT Large still averages a 44.08% error rate across those four subtasks (Mandarin is scored by character error rate), far above its 3.62% on English LibriSpeech.
5. Scaling: Parameters, Pretraining Hours, and the Leaderboard
The largest model in the SUPERB @ SLT 2022 challenge had about twice the parameters of WavLM-large and scored 394 points lower (1242 - 848). Size buys less than the training recipe does. The cleanest evidence is WavLM Base+, which matched its Base size at 94.70M parameters, added more varied training audio, and beat a model more than 3 times its size (316.61 / 94.70 = 3.34).
The challenge score pins the FBANK baseline at 0 and the state of the art of October 15, 2021 at 1000, so anything above 1000 beats that snapshot. The overall score in the WavLM paper is a separate scale, with FBANK at 43.2, and the two must not be compared. Meanwhile pretraining data has grown from the 94,000 hours of WavLM to millions of hours, and the bill now shows up in GPU time.
| Metric | Value | Source |
|---|---|---|
| Overall SUPERB score, HuBERT Large vs WavLM Large (both about 316.6M parameters) | 72.2 vs 74.6 | Microsoft, WavLM, IEEE JSTSP 2022 |
| Overall SUPERB score, WavLM Base+ (94.70M parameters) vs HuBERT Large (316.61M parameters) | 73.4 vs 72.2 | Microsoft, WavLM, IEEE JSTSP 2022 |
| WavLM pretraining audio | 94,000 hours (60,000 Libri-Light, 10,000 GigaSpeech, 24,000 VoxPopuli) | Microsoft, WavLM, IEEE JSTSP 2022 |
| Hidden-set challenge score of WavLM-large, with size and compute | 1242 (315.45M parameters, 4,326G MACs) | Feng et al., SUPERB @ SLT 2022 |
| Hidden-set challenge score of the largest submission, Unsupervised ASR + T5 (619.82M parameters) | 848 | Feng et al., SUPERB @ SLT 2022 |
| XLS-R pretraining audio across five datasets | 436,000 hours | Mohamed et al., IEEE JSTSP 2022 |
| XEUS model size and pretraining compute | 577M parameters; over 1 million hours of audio; 63,000 GPU hours on 64 A100 GPUs over about 40 days | Chen et al., XEUS, EMNLP 2024 |
| Google Universal Speech Model pretraining audio | 12 million hours across over 300 languages | Google, USM paper 2023 |
MACs (multiply-accumulate operations) estimate how much compute a model needs to run. The Google row is a scale reference only: we found no SUPERB score for that model.
6. Multilingual Leaderboards: Does the English Ranking Travel?
Only partly. WavLM, a leader on English SUPERB, was the weakest baseline in the 2025 multilingual challenge, with 72.0% character error rate against 40.2% for XEUS. The challenge paper describes that model as English-only, which suggests a leaderboard rank says as much about the pretraining audio as about the architecture. For how speech encoders behave across languages at the embedding level, see our multilingual speech embedding statistics.
ML-SUPERB rescales its score the same way, with FBANK at 0. Scaling also backfires here: large models usually won on single-language ASR, but the authors found base models tend to generalize better to multilingual cases than their large versions. The ML-SUPERB 2.0 challenge then shows dialects as the unsolved part, with the XEUS character error rate rising from 40.2% to 68.7%.
| Metric | Value | Source |
|---|---|---|
| ML-SUPERB training audio across 143 languages, 10-minute set vs 1-hour set | 37.43 hours vs 222.46 hours | Shi et al., ML-SUPERB, Interspeech 2023 |
| ML-SUPERB score of XLSR-128, the top model, on the 10-minute training set vs the 1-hour training set | 947.5 vs 996.0 | Shi et al., ML-SUPERB, Interspeech 2023 |
| ML-SUPERB score at 10 minutes, wav2vec2-base vs wav2vec2-large | 755.2 vs 598.3 | Shi et al., ML-SUPERB, Interspeech 2023 |
| ML-SUPERB score at 10 minutes / 1 hour: XEUS (577M parameters, 1M hours) vs MMS 1B (1B, 491K hours) vs w2v-BERT 2.0 v2 (580M, 4.5M hours) | 956 / 956 vs 953 / 948 vs 826 / 916 | Chen et al., XEUS, EMNLP 2024 |
| Character error rate on standard speech, XEUS vs English-only WavLM | 40.2% vs 72.0% | Chen et al., ML-SUPERB 2.0 Challenge, Interspeech 2025 |
| Character error rate of XEUS, standard speech vs dialectal speech | 40.2% vs 68.7% | Chen et al., ML-SUPERB 2.0 Challenge, Interspeech 2025 |
w2v-BERT 2.0 v2 had the most pretraining audio in that comparison and the lowest score, so more hours alone did not settle it either. Scores from the original ML-SUPERB and the XEUS comparison should be read within each paper, not across them.
7. Efficiency and Real-World Uptake
The leading models are expensive, and a line of research now tries to shrink them. DistilHuBERT is the clearest example, but the challenge data shows the price: at 27.03M parameters it scored 617, against 784 for HuBERT-base at 94.70M. The smallest challenge model scored just 104, so compression still loses more than the headline percentages imply.
Benchmarking itself has been a cost too, which is what MiniSUPERB targets. Download counts add a more recent signal: wav2vec2-base-960h out-pulls WavLM Large and HuBERT Large even though the leaderboard placed it lower. Being easy to adopt, and being first, count for a lot.
| Metric | Value | Source |
|---|---|---|
| DistilHuBERT compared with HuBERT | 75% smaller, 73% faster | Chang et al., DistilHuBERT, ICASSP 2022 |
| Hidden-set challenge score, DistilHuBERT (27.03M parameters) vs HuBERT-base (94.70M parameters) | 617 vs 784 | Feng et al., SUPERB @ SLT 2022 |
| Smallest challenge model, Chimera MelHuBERT v1 | 22.50M parameters, 271G MACs, score 104 | Feng et al., SUPERB @ SLT 2022 |
| MiniSUPERB compute reduction, and its rank correlation with the SUPERB challenge | 97% fewer MACs; Spearman 0.982 | Wang et al., MiniSUPERB, IEEE ASRU 2023 |
| GPUs needed: SUPERB recipes vs official wav2vec 2.0 Large fine-tuning recipe | 1 GPU vs around 4 to 8 GPUs | Yang et al., IEEE/ACM TASLP 2024 |
| Downloads last month, facebook/wav2vec2-base-960h | 1,737,874 | Hugging Face model page, October 31, 2026 |
| Downloads last month, microsoft/wavlm-large | 1,068,711 | Hugging Face model page, October 31, 2026 |
| Downloads last month, facebook/hubert-large-ls960-ft | 616,230 | Hugging Face model page, October 31, 2026 |
A download counter records pulls, not people, so read these rows as a sign of activity rather than a count of users.
Summary: Self-Supervised Speech Model SUPERB Leaderboard by the Numbers
| Metric | Value | Source |
|---|---|---|
| WER with 10 minutes of labels, wav2vec 2.0 Large (test-clean / test-other) | 4.8% / 8.2% | Meta AI, wav2vec 2.0, NeurIPS 2020 |
| WER with all 960 hours of labels, wav2vec 2.0 Large (test-clean / test-other) | 1.8% / 3.3% | Meta AI, wav2vec 2.0, NeurIPS 2020 |
| Phoneme error rate, FBANK vs HuBERT Large | 82.01% vs 3.53% | Yang et al., SUPERB, Interspeech 2021 |
| Speech recognition WER on SUPERB, WavLM Large vs XEUS | 3.44% vs 3.34% | Chen et al., XEUS, EMNLP 2024 |
| Emotion recognition accuracy on IEMOCAP, XEUS | 71.08% | Chen et al., XEUS, EMNLP 2024 |
| Speaker verification EER on VoxCeleb1-O, ECAPA-TDNN reference vs fine-tuned WavLM Large | 1.010% vs 0.383% | Microsoft, WavLM, IEEE JSTSP 2022 |
| Speech translation BLEU, FBANK vs HuBERT Large | 2.32 vs 20.01 | Tsai et al., SUPERB-SG, ACL 2022 |
| Models beating FBANK on all 15 tasks | About half of 33 | Yang et al., IEEE/ACM TASLP 2024 |
| Hidden-set challenge score of WavLM-large | 1242 | Feng et al., SUPERB @ SLT 2022 |
| Overall SUPERB score, WavLM Base+ vs HuBERT Large | 73.4 vs 72.2 | Microsoft, WavLM, IEEE JSTSP 2022 |
| XEUS pretraining compute | 63,000 GPU hours | Chen et al., XEUS, EMNLP 2024 |
| Google Universal Speech Model pretraining audio | 12 million hours | Google, USM paper 2023 |
| ML-SUPERB score of XLSR-128 on the 1-hour training set | 996.0 | Shi et al., ML-SUPERB, Interspeech 2023 |
| Character error rate on standard speech, XEUS vs English-only WavLM | 40.2% vs 72.0% | Chen et al., ML-SUPERB 2.0 Challenge, Interspeech 2025 |
| DistilHuBERT compared with HuBERT | 75% smaller, 73% faster | Chang et al., DistilHuBERT, ICASSP 2022 |
| MiniSUPERB compute reduction | 97% fewer MACs | Wang et al., MiniSUPERB, IEEE ASRU 2023 |
| Downloads last month, facebook/wav2vec2-base-960h | 1,737,874 | Hugging Face model page, October 31, 2026 |
Methodology and Sources
- SUPERB: Speech processing Universal PERformance Benchmark (Yang et al., Interspeech 2021)
- SUPERB-SG (Tsai et al., ACL 2022)
- A Large-Scale Evaluation of Speech Foundation Models (Yang et al., IEEE/ACM TASLP 2024)
- SUPERB @ SLT 2022 challenge report (Feng et al., 2022)
- ML-SUPERB (Shi et al., Interspeech 2023)
- The ML-SUPERB 2.0 Challenge (Chen et al., Interspeech 2025)
- MiniSUPERB (Wang et al., IEEE ASRU 2023)
- wav2vec 2.0 (Meta AI, NeurIPS 2020)
- HuBERT (Meta AI, IEEE/ACM TASLP 2021)
- wav2vec-U (Meta AI, 2021)
- WavLM (Microsoft, IEEE JSTSP 2022)
- XEUS: Towards Robust Speech Representation Learning for Thousands of Languages (Chen et al., EMNLP 2024)
- Universal Speech Model (Google, 2023)
- DistilHuBERT (Chang et al., ICASSP 2022)
- Self-Supervised Speech Representation Learning: A Review (Mohamed et al., IEEE JSTSP 2022)
- Hugging Face model pages: facebook/wav2vec2-base-960h, microsoft/wavlm-large, facebook/hubert-large-ls960-ft (October 31, 2026)
- Data watch: the live SUPERB leaderboard could not be read on October 31, 2026 (the host refused the connection and the benchmark’s GitHub page shows no scores), so every SUPERB figure here comes from the published papers and a newer submission may exist. Three score scales appear and are not interchangeable: the overall score in the WavLM paper, the challenge score (baseline 0, October 15, 2021 state of the art 1000), and the ML-SUPERB score. The 586.9 score for wav2vec2-large on the ML-SUPERB 1-hour set follows a printed language identification value of 8.0 in that paper’s table, far below its neighbors, so that single gap should be read with care. Parameter counts vary slightly by paper: 316.62M vs 315.45M for WavLM Large and 94.70M vs 94.38M for WavLM Base+. Minor conflicts inside sources: the SUPERB text says 5.10 for HuBERT speaker verification while its table says 5.11 (we use the table); the XEUS text gives 33.3 for its best 10-minute character error rate while its table gives 30.3 for XEUS and 33.3 for MMS 1B; the ML-SUPERB 2.0 paper states both 149 and 154 languages, and its abstract gains do not match its section 4.1, so we report only its baseline table. Hugging Face downloads are a rolling count of the last month and change daily. Most SUPERB results date from 2021 to 2024, so treat them as most recent available data rather than 2026 measurements.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the next ML-SUPERB challenge results are published.