Skip to main content

NetEase YoudaoAudio modelOpen weights

Confucius4-R2T2

PopularityRank 10 of 24 audio models

39out of 100new calculation

based on two data groups, mostly search, range 29 to 49

Summary

Confucius4-R2T2 is a streaming ASR model for Chinese and English. It builds on Qwen3-ASR-1.7B, keeps an append-only stable transcript prefix, and allows decoding chunks from 80 milliseconds to two seconds.

The model card reports roughly 200 to 600 milliseconds of average latency. The weights run locally under the NetEase Model Use License Agreement.

GraduallyBench Audio

Confucius4-R2T2

Partially tested
  • The test coverage is incomplete.
  • Only a technical test is available.
  • The selection of test inputs is limited.
  • The route does not confirm the exact model version.
  • Silence was not reliably evaluated.
  • Completion in the user interface was not confirmed.
Show test details

Speech recognition

Public demo · German, English

30 valid outputs from 35 attempted and 36 planned test slots.

Word error rate (WER)
11.2 %(58 / 518 reference words)German
Character error rate (CER)
3.77 %(116 / 3,074 reference characters)German
Word error rate (WER)
2.12 %(10 / 471 reference words)English
Character error rate (CER)
1.17 %(29 / 2,481 reference characters)English
Word error rate (WER)
6.88 %(68 / 989 reference words)German, English
Character error rate (CER)
2.61 %(145 / 5,555 reference characters)German, English
Silence assessment
Not availableGerman, English

Technical measurement

Public demo · English

1 valid output from 1 attempted and 1 planned test slots.

  • Only a short functional test is available.
  • The functional test does not assess content quality.
  • Only the indicated language subset was tested.

No further scored metric is available for this test.

WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 44 published benchmark results

Page 1 of 4

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Confucius4-R2T2 Quality, AISHELL-1 CER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
2.29
Metric
character error rate (CER) on AISHELL-1 test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on AISHELL-1. The authors report character-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, AISHELL-2 CER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
3.58
Metric
character error rate (CER) on AISHELL-2 test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on AISHELL-2. The authors report character-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, AISHELL-3 CER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
2.85
Metric
character error rate (CER) on AISHELL-3 test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on AISHELL-3. The authors report character-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, WenetSpeech Meeting CER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
7.37
Metric
character error rate (CER) on WenetSpeech Meeting test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on WenetSpeech Meeting. The authors report character-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, WenetSpeech Net CER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
8.48
Metric
character error rate (CER) on WenetSpeech Net test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on WenetSpeech Net. The authors report character-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, LibriSpeech test-clean WER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
2.11
Metric
word error rate (WER) on LibriSpeech test-clean test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on LibriSpeech test-clean. The authors report word-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, LibriSpeech test-other WER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
4.81
Metric
word error rate (WER) on LibriSpeech test-other test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on LibriSpeech test-other. The authors report word-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, GigaSpeech clean WER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
10.21
Metric
word error rate (WER) on GigaSpeech clean test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on GigaSpeech clean. The authors report word-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, VoxPopuli WER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
7.26
Metric
word error rate (WER) on VoxPopuli test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on VoxPopuli. The authors report word-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
Confucius4-R2T2 Quality, TED-LIUM WER (X2Streaming-ASR Table 1) (Show measurement, test conditions, and source)
Source value
3.45
Metric
word error rate (WER) on TED-LIUM test set, streaming Quality configuration
Unit
%
Direction
Lower is better
Source type
independent-evaluator
Evaluator
X2Streaming-ASR paper, X Square Robot
Status
active
Retrieved at
2026-10-22
Methodology
Independent evaluation. Confucius4-R2T2 in the streaming inference setting “Quality” on TED-LIUM. The authors report word-level error rate and emission latency relative to forced-aligned reference endpoints.
Context
The paper does not state an exact Confucius checkpoint revision, per-test-set sample count, confidence interval, or measurement date. The Quality column is distinct from the Speed setting and the shared Offline results.
Source X2Streaming-ASR paper, X Square Robot
AMI WER (Show measurement, test conditions, and source)
Source value
11.37
Metric
word error rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
Confucius4-R2T2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. 160 ms representative streaming configuration; true streaming, append-only
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Confucius4-R2T2 model card
GigaSpeech clean WER (Show measurement, test conditions, and source)
Source value
9.6
Metric
word error rate
Unit
%
Benchmark version
source version not stated
Category
audio
Direction
Lower is better
Source type
vendor-reported
Evaluator
Confucius4-R2T2 model card
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. 160 ms representative streaming configuration; true streaming, append-only
Context
Source-specific observation; it does not establish a compatible cross-model cohort.
Source Confucius4-R2T2 model card

Page 1 of 4

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 54 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

Confucius4-R2T2Eleven v3 39GPT-Live-Transcribe 39Scribe v2 39Sonic 3.6 39

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Model class
Streaming speech recognition
Source: Model class
Checkpoint
netease-youdao/Confucius4-R2T2
Source: Checkpoint
Base model
Built on Qwen/Qwen3-ASR-1.7B
Source: Base model

Input and output

Input
Chinese and English audio
Source: Input
Output
Stable streaming transcript
Source: Output
Latency
About 200 to 600 milliseconds on average
Source: Latency

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Access
Local weights
Source: Access
License
NetEase Model Use License Agreement
Source: License
Shutdown
None announced
no source

Further details

Decoding
Configurable chunks from 80 milliseconds to 2 seconds
Source: Decoding

Sources and data date

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.