GREG How-To Videos – Voicegain

Here are YouTube links to videos showing aspects of working with GREG:

Uploading existing data

This is relevant if you already possess audio and recognition results and want to load them to GREG. Normally such data would be collected live from a suitably configured IVR application that uses GRXML. Each question would be generating data for a separate GREG Experiment. (We will be expanding the Telephony Bot API to also be able to send recognition results to GREG.)

Here is an example input file for an Experiment where we recognize sequences of 60 digits (audio files are named: 1.wav, 2.wav, 3.wav, etc).

name result utterance confidence interpretation
1 MATCH "9 9 9 8 1 5 9 3 4 1 5 5 0 4 3 1 0 8 3 9 8 2 9 9 6 7 2 5 7 7 3 3 8 5 6 5 1 7 1 6 4 5 3 4 4 1 0 2 2 7 0 3 5 2 8 5 8 1 7 3" 0.9 digits="999815934155043108398299672577338565171645344102270352858173"
2 MATCH "5 4 7 8 6 0 7 9 9 6 8 9 6 7 6 9 0 6 1 0 6 4 3 0 0 3 1 6 5 6 1 7 4 3 5 9 8 8 9 4 9 0 0 8 9 8 2 4 9 6 5 2 3 4 1 7 7 6 4 2" 0.9 digits="547860799689676906106430031656174359889490089824965234177642"
3 MATCH "2 2 8 0 1 6 7 4 4 5 4 3 5 9 1 1 2 3 3 6 6 4 5 4 1 0 8 7 9 9 1 8 5 7 7 3 5 1 0 8 1 3 5 3 4 5 5 6 8 8 4 7 3 2 4 5 8 6 5 0" 0.9 digits="228016744543591123366454108799185773510813534556884732458650"
4 MATCH "3 9 0 5 0 8 3 1 9 5 0 7 8 3 2 7 7 4 8 3 7 7 6 0 3 0 5 4 2 4 3 9 2 5 1 9 5 1 3 3 4 0 3 1 3 5 3 0 5 4 4 3 1 3 7 6 5 9 4 8" 0.9 digits="390508319507832774837760305424392519513340313530544313765948"
5 MATCH "8 1 1 2 0 7 7 5 1 0 1 0 2 3 6 1 8 8 4 7 5 8 5 6 4 8 7 1 1 2 6 3 1 3 5 9 3 5 8 5 6 3 6 7 9 4 9 4 5 1 8 5 2 7 5 1 6 2 1 8" 0.9 digits="811207751010236188475856487112631359358563679494518527516218"
6 MATCH "2 7 5 1 5 2 6 8 9 4 8 8 5 1 1 2 5 5 0 9 1 8 8 1 6 8 6 5 4 0 1 4 4 0 4 0 6 7 8 6 5 9 0 2 7 2 6 4 5 4 7 6 5 8 4 2 2 8 2 8" 0.9 digits="275152689488511255091881686540144040678659027264547658422828"
7 MATCH "6 2 8 4 7 2 1 2 2 4 0 9 2 9 3 8 2 4 1 5 5 1 6 0 5 3 7 9 9 5 0 6 3 6 6 3 3 4 2 3 8 1 3 6 1 3 5 8 2 5 8 6 9 6 4 2 9 2 7 8" 0.9 digits="628472122409293824155160537995063663342381361358258696429278"
8 MATCH "8 7 0 3 1 4 1 2 2 8 8 6 8 3 4 1 2 0 7 7 8 0 0 8 9 8 5 0 9 9 2 9 7 5 2 9 6 2 4 8 0 9 5 4 5 9 0 0 3 4 8 1 1 5 0 5 4 3 8 9" 0.9 digits="870314122886834120778008985099297529624809545900348115054389"
9 MATCH "0 3 0 1 3 0 1 4 9 2 2 9 2 6 9 5 5 3 2 3 9 9 2 1 6 1 6 4 2 3 4 9 0 3 4 0 5 6 7 6 6 5 9 3 6 1 5 6 3 7 5 7 6 3 3 5 2 3 1 1" 0.9 digits="030130149229269553239921616423490340567665936156375763352311"
10 MATCH "9 1 5 4 2 0 9 9 9 9 9 3 7 2 3 8 5 7 7 5 5 4 5 9 7 9 1 1 5 0 5 2 5 7 7 6 3 4 0 9 8 6 9 5 6 0 9 7 1 5 6 3 9 9 5 5 6 1 7 9" 0.9 digits="915420999993723857755459791150525776340986956097156399556179"

Preparing files to upload experiment data to GREG (4:10) - this does not show how to setup question and grammar
Upload Experiment data to GREG (3:26) - this does not show how to prepare the upload data

Analyzing data using GREG

Reviewing a data set in GREG (5:41) - it shows how to:
- enter true utterance and interpretation for each utterance, including cases of NOMATCH
- notice that NOMATCH is in our system considered the same as a special __garbage__ interpretation returned from the grammar
Running GREG Experiment (4:32) - We demonstrate running two GREG Experiments, each with a different Acoustic Model, and then comparing the results using GREG browser. Note: this video was taken in back in June 2020, the accuracy of the current Voicegain model, specifically on number sequences, is significantly higher than what can be seen in the video.
- This demo changes the NN model between the experiments, but you can also modify the grammar, which likely would be more common, see dialog at time 2:51.