Takashi Okura

Hello everyone!

In a previous article I wrote titled "Why speech recognition rates are not 100%?", I mentioned that one of the reasons is thatVariations in spelling" was introduced.

Why is speech recognition never 100%?

AmiVoice Cloud Platform-Tech Blog

In this article, we will delve a little deeper into this "spelling variation." In addition to explaining what spelling variation is,

• In what situations does the variation in spelling become a problem?
• How much orthographic variation occurs in speech recognition?

I would also like to explain these points.

Table of contents

• What is spelling variation?
• When spelling variations are a problem
  • Text search becomes difficult
  • May affect voice recognition rate
  • Format mismatch
• To what extent does spelling variation occur?
• Final thoughts

What is spelling variation?

"Spelling variation" refers to the fact that words have the same sound and meaning but are written differently. For example, there are the following patterns:

Pattern	An example
Kanji and Kana	various / various
Differences in okurigana	Moving / Moving / Moving
Differences in kana spelling	Violin
Alphabet and Kana	AmiVoice
Kanji and Arabic numerals	2000 yen / 2000 yen

Furthermore, Ando-san, who often appears in this blog,King of Words with Variations in Spelling" is what is being said.Seeker". Just thinking about it.

• Seeker
• Citrus depressa
• Shikuwasa
• Shiikuwasha
• ...

And there's no end to it.*1.

When spelling variations are a problem

In what cases does this "variation in spelling" cause problems? Here are three examples of problems.

Text search becomes difficult

I think the most common way to utilize the results of speech recognition is to "convert the speech into text and save it" like a meeting minutes. One of the benefits of this is thatYou can search for the content later"That's it.

Let's consider the case where there is variation in the text of the recognition results. For example, if you search for "AmiVoice" in the text, "AmiVoice" will not appear in the search results if the condition is an exact match.

Using OR searches or regular expressions, you can deal with slight variations in spelling, but when it comes to something like "shikwasa," it becomes difficult to cover all patterns.*2.

May affect voice recognition rate

As I mentioned in a previous article, speech recognition rates are basically calculated on a character-by-character basis, so variations in spelling can sometimes be counted as errors.

Regarding the "recognition accuracy (recognition rate)" of speech recognition

AmiVoice Cloud Platform-Tech Blog

For example, the correct sentence is "move" but the recognition result is "move."っYueTooth", if it is left as it is, it will become "っ"When"Tooth" partInsertion errorThis will count as a lower apparent voice recognition rate.

We will take a closer look at the extent to which spelling variations occur and how they affect speech recognition rates in the next section.

Format mismatch

This is often the case with matters such as parliamentary minutes, where the format of numbers and other characters is strictly determined. For example, if the rule is that "numbers should be displayed in Arabic numerals," but the recognition result outputs "two thousand yen," this is unacceptable even if it is a variation in notation. Measures such as reviewing the training data and converting the Chinese numerals in the recognition results to Arabic numerals in post-processing are required.

To what extent does spelling variation occur?

How much variation in spelling occurs in actual speech recognition?

This is a verification using internal data,Differences in okurigana""The difference between kanji and hiragana*3" and other variations in spelling, the recognition rateAbout 1 to 2 pointsIt seems to rise*4Of course, this value will vary depending on the audio situation and the engine, so it is only a reference value.

There were also cases like this:
We previously published an article comparing the speech recognition rates (recognition accuracy) of "general-purpose engines" and "domain-specific engines."

Comparing the speech recognition accuracy of AmiVoice's domain-specific engines (General-prupose vs Electronic Medical Records)

AmiVoice Cloud Platform-Tech Blog

The speech recognition rate shown in this article is

Misunderstandings caused by spelling variations have been corrected by revising the answer sentences.

As stated, these are the values ​​after correcting the variations in notation.

Engine	Voice recognition rate
Voice Input_General-Purpose	87.41%
Voice Input_Electronic Medical Record	97.61%

The table below also shows the speech recognition rates of these two engines before any spelling variations were corrected.

Engine	Voice recognition rate (Before modification → After modification)
Voice Input_General-Purpose	75.73% → 87.41%
Voice Input_Electronic Medical Record	93.77% → 97.61%

What do you think? The general-purpose engine's recognition rate has improved by more than 10 points due to the correction of the notation variations, and some people may be surprised at the difference, which is much larger than they thought. However, when I checked the recognition results, I got the following message:Difference between full-width and half-width""The difference between Arabic numerals and Chinese numerals" There were many spelling variations related to formatting differences in the test set. If you are not getting the recognition rate you expect with speech recognition, we recommend checking the recognition results and the correct sentences to see if there are any spelling variations, including formatting differences.

Final thoughts

In this article, we have looked in detail at "orthographic variation" in speech recognition. As mentioned in the article "Why is speech recognition never 100% accurate?", there are many orthographic variations, and in many cases they cannot be converted mechanically. We hope that you can see that this is one of the difficult issues when dealing with speech recognition.

If any developers have become interested in voice recognition technology or the AmiVoice Cloud Platform after reading this article, please feel free to contact us. https://acp.amivoice.com/ Give it a try. You can use all engines for free up to 60 minutes of audio per month.

Thank you for reading this far!