Hello everyone.

This is the final installment of our series on tips for improving the accuracy of speech recognition. The final topic will be "the appropriate communication environment for speech recognition."

Because speech recognition consumes a large amount of resources, such as CPU/GPU and memory, it is often processed on a cloud server rather than on a PC or smartphone. This makes the network communication environment important.

This time, we will explain in detail the network and wireless communication environment such as Bluetooth when using a speech recognition server.

An "appropriate communication environment" is essential for speech recognition!

First, we will explain the amount of communication data required for speech recognition.

The above is the main data exchanged when using the AmiVoice API, but other speech recognition engines may send and receive other information.

The key point is that the majority of communication data consists of audio data transmitted from the client to the server. Other than audio data, the information is primarily text-based, and typically the data volume is not large.

For speech recognition, a sampling rate of 16 kHz and a quantization bit depth of 16 bits are common, and the resulting data volume per second is calculated as follows:

• 16 kHz × 16 bit = 256 kbit/s = 32 kB/s

Please note that the amount of audio data will vary depending on the following factors:

• When processing multiple channels of sound simultaneously (for example, in the case of two channels, the amount of data is twice that of one channel)
• When transmitting higher quality audio data (for example, if the sampling rate is 24 kHz, the amount of data is 1.5 times that of 16 kHz)
• When transmitting low-quality audio data, such as landline voice (the sampling rate is often set to 8 kHz, which is half the data volume of 16 kHz)
• When compression formats are supported (the data volume decreases according to the compression ratio)

Data other than audio data is usually not that large, but under certain conditions, large amounts of data may be sent or received. For example, with the AmiVoice API, it is possible to send a large word registration list or grammar definitions (rule grammar) for speech recognition when making a request. Please estimate the size of this data as needed.

With this in mind, it is necessary to have an appropriate communication environment and sufficient bandwidth when using speech recognition. If the communication environment is poor and the bandwidth is insufficient, the following two issues may occur:

• For streaming
  The sound transmission may be delayed, or may even be dropped, or may result in a timeout error.
• When sending audio files
  It takes time to send the voice, which increases the time it takes for the speech recognition results to be returned, and in some cases it may cause a timeout error.

Additionally, apart from network communication, speech recognition accuracy may decrease when the communication environment for wireless microphones or Bluetooth microphones is poor. Possible causes for this include crosstalk, radio wave interference, and obstacles. If sound is interrupted or voice is not transmitted correctly, please check the communication environment between the microphone and the receiving device and attempt to improve it.

What to do if the communication environment is poor

Here are some ways to deal with poor communication conditions.
First, if the communication environment is unstable, there is a method to build a speech recognition server within the local network instead of using a cloud server on the Internet. This makes it possible to reduce or eliminate network latency and instability. However, the speech recognition engines that support this measure are limited. In the case of AmiVoice, this is provided through a plan called "AmiVoice API Private." Additionally, for example, by using OpenAI's "Whisper," it is also possible to build your own server.

Next, there is also a method to perform speech recognition on a PC or smartphone without using a server. In this case, even a local network is not required, and speech recognition can be executed completely offline. In the case of AmiVoice, this offline speech recognition is possible by using "AmiVoice SDK."

When the network communication environment is poor
- Building a speech recognition server within a local network
▶ Available in AmiVoice API Private
▶ There are engines such as Whisper that allow you to build your own server
- Offline speech recognition within the device
▶ Possible with AmiVoice SDK

Next, we will explain what to do if the communication environment for wireless microphones or Bluetooth microphones is poor.

First of all, the most reliable method is to use a wired microphone or the microphone built into the device whenever possible. However, if you need to use a wireless or Bluetooth microphone, the following measures can be considered.

• Avoiding radio interference
  It is important to avoid interference by moving away from other wireless/Bluetooth devices, microwave ovens, and other devices that emit radio waves. Since interference is more likely to occur in crowded places, it is also effective to move away from such environments.
• Reduce the impact of obstacles
  Obstacles such as walls and furniture can increase the chance of communication being interrupted. You can improve the communication environment by reducing obstacles or by moving the device's transmitter and receiver closer together.
• Change the radio channel
  Depending on the device you are using, you may be able to change the radio channel, which may allow you to switch to a channel that does not interfere with other devices.
• Contact the manufacturer
  Depending on the microphone, there may be other measures available besides those listed above, so it may be a good idea to consider contacting the microphone manufacturer for detailed instructions.

When the communication environment for wireless microphones or Bluetooth microphones is poor
- Use a wired or on-device microphone
- Stay away from environments with radio interference
- Reduce obstacles or shorten the distance between the transmitter and receiver
- Change the radio channel you use
- For details, contact the microphone manufacturer

Summary of tips to improve the accuracy of speech recognition

So far, we have discussed tips for improving the accuracy of speech recognition in five articles. Let's summarize the key points once again.

Points that influence the accuracy of speech recognition
① Speak appropriately
② Select the right device
③ Use the device appropriately
④ Use in a low-noise environment
⑤ Use in an appropriate communication environment

We hope that you will put into practice the points introduced here and use your ingenuity to make the most of speech recognition in a better environment.