Model
* propose and support U2++, as the following graph shows, which uses both forward and backward information at training and decoding.
* support dynamic left chunk training and decoding, so we can limit history chunk at decoding to save memory and computation.
* support distributed training.
Dataset
Now we support the following five standard speech datasets, and we got SOTA result or close to SOTA result.
| 数据集 | 语言 | 数据量(h) | 测试集 | CER/WER | SOTA |
|-------------|------|-----------|------------|---------|---------------|
| aishell-1 | 中文 | 200 | test | 4.36 | 4.36(WeNet) |
| aishell-2 | 中文 | 1000 | test_ios | 5.39 | 5.39(WeNet) |
| multi-cn | 中文 | 2385 | / | / | / |
| librispeech | 英文 | 1000 | test_clean | 2.66 | 2.10(EspNet) |
| gigaspeech | 英文 | 10000 | test | 11.0 | 10.80(EspNet) |
Productivity
Here are some features related to productivity.
* LM support. Here is the system design or LM supporting. WeNet can work with/without LM according to your applications/scenarios.
* timestamp support.
* n-best support.
* endpoint support.
* gRPC support
* further refine x86 server and on-device android recipe.