TeleFlow SpeechTrainer
From TeleFlow
TeleFlow SpeechTrainer allows you to review utterances from users of a Speech Recognition (SR) system and make adjustments to the grammars so the user experience gradually improves. SpeechTrainer helps you to determine what happened with all of the SR recognitions on a call by call basis. Then you can make adjustments as needed to how you prompt your caller, and how information is requested from the caller.
Speech training is a process comprising many things, such as modifying grammars, the way in which the caller is prompted, adding confirmations where callers are having trouble, etc.
Please refer to the planning and developing a speech recognition system for considerations before you start building a speech recognition environment.
Contents |
Setting Up the TeleFlow to Log Speech Recognition Events
Run TFConfig and check on “Record Speech Recs” click "Save SR Audio to Log Directory"
Restart TeleFlow Server to accept the new setting. Each day a new folder is created of Speech Recognitions and related files. These are files that are created when callers are speaking to provide voice commands to the recognizer (ie: Nuance SR / LumenVox SR)
Transferring Data to the SpeechTrainer Computer
SR files will create on a daily basis. At the start of the day, the previous day’s SR should be reviewed. You can review the files on the TeleFlow Server computer if it has a sound card. Generally however, SR files will be reviewed on another desktop computer.
Log files are:
- Generated once per day per port
- Contains information about only those recognition results flagged by the application via the SR Flag Result step (refer to Designer for this step).
- Contains comma separated data, just like a .csv file.
- Location:{log}\ ST{date}\
where:
- {log} = TeleFlow log directory
- {date} = The current date yyyymmdd
- File name:ST{date}-{port}.FLG:
where:
- {date} = The current date yyyymmdd
- {port} = The port number
To bring back those files, copy the needed log file (.Log) into the that days folder as described above, and copy the entire directory (folder) to the client computer.
Using SpeechTrainer Interface
Install the SpeechTrainer software on the computer you will be using to review the SR output by. This program is a separate install, and must be installed on a computer that is running Windows XP, 2000 or 2003 with a sound card.
- Start SpeechTrainer
- Click the “Open CSV” button, and choose the SR.csv from the directory
- The left hand pane is each complete call that was made. The right hand pane shows the speech recognitions that were recorded.
When you click on a line in the SR pane the details for that SR are detailed below it.
If the previous play to the SR is the same as the last heard play then only the previous play details are populated.
When you click on the flag button, a copy of that SR is placed in the SR dir\Flag for later use. Also the number row of the SR is marked yellow.
If you press play for the "Previous Play" or "Last Heard Play" and the file does not exist to the absolute path, then a dialog will open. Once you choose the location of the file, every other play will attempt to look in that directory as well.
Show log will open the log file, as long as it has been moved into the SR directory. Larger files can take a few seconds to open. Once the log file is up, you can press the find button. This find dialog will be populated with the name of the SR.wav.
Display failed Speech Recs Only: will only display recognitions where the SR Result Confidence is less than Threshold Confidence.
Improving SR Applications with SpeechTrainer
The SpeechTrainer will help with several areas. Determining what the caller is asking for at which point in the system, and pronunciation are the primary benefits of using SpeechTrainer.
SpeechTrainer will allow you to listen to the responses (or utterances) from the callers. By reviewing the utterances, you can determine what the expectations of the caller were, or it they were speaking commands that were not accounted for.
Grammars can then be adjusted and provided back to TeleFlow Server for the next days testing or usage.
SpeechTrainer will also help determine the style of the application. For example, occasionally a critical piece of information is requested by the system. During these points you may need to restate what was heard so that you can be certain. You don't need confirmations everywhere, but you can use them for specific areas you are concerned with, or where difficulties arise. So, if in speech training, you do find that "Option A" and "Option B" don't work as well as you need them to, confirm them with the caller, like this: "Did you say Option A?" and retrieve the "Yes" or "No" response, which has very distinct and dependable results.
SpeechTrainer will not specifically indicate how to solve a problem, but will provide insight into where callers are having trouble.
New systems should be reviewed on a daily basis in order to improve grammars. Overtime, reviews can be reduced to once a week once stabilized.
