Model card
Google USM.
Googleproprietary2B paramsConformer encoder + RNN-T/CTC
Universal Speech Model. Pretrained on 12M hours unlabeled + 28M hours supervised. 100+ languages.
§ 01 · Benchmarks
Every benchmark Google USM has a recorded score for.
| # | Benchmark | Area · Task | Metric | Value | Rank | Date | Source |
|---|---|---|---|---|---|---|---|
| 01 | LibriSpeech | Speech · Speech Recognition | wer-test-clean | 2.0% | #2 | 2023-03-02 | source ↗ |
| 02 | LibriSpeech | Speech · Speech Recognition | wer-test-other | 4.1% | #3 | 2023-03-02 | source ↗ |
Rank column shows this model’s position vs all other models scored on the same benchmark + metric (competitors after the slash). #1 in red means current SOTA. Sorted by rank, then newest result.
§ 03 · Papers
1 paper with results for Google USM.
- 2023-03-02· Speech· 2 results
Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages
§ 04 · Related models
Other Google models scored on Codesota.
Gemini 2.5 Pro
16 results · 3 SOTA
Gemini 3 Pro
Undisclosed params · 13 results · 2 SOTA
Gemini 1.5 Pro
12 results · 1 SOTA
Gemini 3.1 Pro
3 results · 1 SOTA
ViT-H/14
632M params · 2 results · 1 SOTA
CoCa (finetuned)
2.1B params · 1 result · 1 SOTA
Gemini 2.0 Flash
1 result · 1 SOTA
Gemini 3.1 Pro Preview
1 result · 1 SOTA
§ 05 · Sources & freshness
Where these numbers come from.
arxiv
2
results
2 of 2 rows marked verified.