Multi_zh-Hans Recipe

[JinZr]

This PR includes scripts for training Zipformer model using multiple Chinese datasets.

Included Training Sets

1. THCHS-30
2. AiShell-{1,2,4}
3. ST-CMDS
4. Primewords
5. MagicData
6. Aidatatang_200zh
7. AliMeeting
8. WeNetSpeech
9. KeSpeech-ASR

Included Test Sets

1. Aishell-{1,2,4}
2. Aidatatang_200zh
3. AliMeeting
4. MagicData
5. KeSpeech-ASR
6. WeNetSpeech

[JinZr]

@csukuangfj all requested changes have been applied, thank you!

[xiaoxi91]

Hi, I was exploring this recipe and noticed that the Chinese modeling unit has been adjusted from char to high-freq-char + byte-symbol. I'm curious about the rationale behind this change and would love to understand the thought process. Could you please shed some light on why this decision was made?

Moreover, I'm wondering if this modification is expected to improve performance in any way? If so, could you elaborate on how and in which specific scenarios this might be beneficial?

Thank you very much for your time and for sharing this valuable recipe! Looking forward to your insights.

[JinZr]

Dear User, thanks for raising the question! the motivation for this modeling unit partitioning protocol is that when we looked into the histogram of characters exported from all corpora involved, we found that the distribution of the characters was highly unbalanced, i.e. a few portions of the characters makes most of the appearances in the training data. of course you can use the vanilla char based modeling solution, but the main concern is that you will eventually have a output layer with an enormous amount of parameters (i couldn’t recall the exact number) and most of them might not be properly trained. Best Regards Jin

…

On Tue, 13 Aug 2024 at 17:04 xiaoxi91 ***@***.***> wrote: Hi, I was exploring this recipe and noticed that the Chinese modeling unit has been adjusted from char to high-freq-char + byte-symbol. I'm curious about the rationale behind this change and would love to understand the thought process. Could you please shed some light on why this decision was made? Moreover, I'm wondering if this modification is expected to improve performance in any way? If so, could you elaborate on how and in which specific scenarios this might be beneficial? Thank you very much for your time and for sharing this valuable recipe! Looking forward to your insights. — Reply to this email directly, view it on GitHub <#1238 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOON42HSM2NV7YD7VOL7AILZRHDYJAVCNFSM6AAAAABMN2EDQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEOBVG4ZTSNRQGQ> . You are receiving this because you modified the open/close state.Message ID: ***@***.***>