Shogo Ando

Hello everyone.

Today I will explain how to handle fillers (unnecessary words) in the AmiVoice API.

What motivated me to write this article

Something like this happened (fiction)

Commentary

AmiVoice API has an automatic removal function for fillers(unnecessary words), such as "えーと" and "あのー".However, as in the manga example above, there may be cases where the filler is removed, which can be a problem.

The advantages and disadvantages of filler removal can be summarized as follows:

• Benefits
  • The speech recognition results are easier to read as unnecessary characters are removed.
• Drawbacks
  • There is a risk that subtle nuances may be lost in the speech recognition results.

By default, they are removed, but by adding keepFillerToken=1 to the d parameter of the request parameter, it becomes possible to output without removing the fillers.
It's mentioned in the manual here.

Request Parameters | AmiVoice API Manual | AmiVoice Cloud Platform

This section explains the parameters to set when requesting speech recognition with the AmiVoice API. Although the transmission method differs between the HTTP and WebSocket interfaces, the parameters that can be set are the same.

acp.amivoice.com

I actually tried it

Let's try it out. There are several ways to use the AmiVoice API, but this time we'll use the one that's best suited for short audio files. I'll try using the Synchronous HTTP Speech Recognition API, which is suited for short audio files.
The basic steps are as described in this article. This explanation is for a Windows 10 environment.

A simple way to try out voice recognition with AmiVoice on an audio file

In the middle of the article, there is a section where you edit a file called run.bat, and you need to change the second line as follows. The key point is that keepFillerToken=1 is specified in the d parameter. This is where you specify the following.

curl -X POST https://acp-api.amivoice.com/v1/recognize -F u=%AppKey% -F d="grammarFileNames=-a-general keepFillerToken=1" -F a=@../../audio/myvoice.mp3 > result.txt

I prepared an audio file (myvoice.mp3) of me speaking the lines from the manga above, and ran the rewritten 'run.bat' to compare the results.

• Speech recognition result when keepFillerToken=1 is added to the d parameter
  • %えー%%えっとー%%あー%%あのー%僕はいいと思うんですけど、%ん%%まー%%ん%%まー%%まー%いいんじゃないでしょうか
• Speech recognition result when nothing is entered in the d parameter
  • 僕はいいと思うんですけど、いいんじゃないでしょうか

You can see that the filler output changes depending on whether the parameter is set or not.

Also, all filler words are wrapped in single-byte % symbols. This allows developers to choose how to handle fillers for each situation where a speech recognition result appears.

• To retain disfluencies: strip only the surrounding % symbols in your code.
• To fully remove disfluencies: delete all content wrapped in % markers via code.

Conclusion

When using the AmiVoice API by default, fillers are automatically removed, but by using the method described above, you can leave the fillers and decide whether or not to remove them depending on the situation. It might be worth considering for better speech recognition results.