Business Context
A banking call center wants to automate the transcription of customer support calls to analyze sentiment and intent. The speech-to-text model requires standard Mel-Frequency Cepstral Coefficients (MFCC) as input. The data pipeline must process raw audio blobs and output a feature matrix retaining the call ID.
Data Preparation
Creation of a simulated table 'CALL_LOGS' containing call IDs and raw binary audio data.
Copied!
| 1 | DATA casuser.call_logs; LENGTH call_id $10 audio_data $2000; INPUT call_id $ audio_data $; DATALINES; |
| 2 | CUST_001 |
| 3 | CUST_002 |
| 4 | ; RUN; |
Étapes de réalisation
1
Execution of feature extraction using standard MFCC parameters (13 coefficients) and transferring the Call ID.
Copied!
| 1 | PROC CAS; |
| 2 | audio.computeFeatures / |
| 3 | TABLE={name='call_logs', caslib='casuser'} |
| 4 | audioColumn='audio_data' |
| 5 | copyVars={'call_id'} |
| 6 | mfccOptions={nCeps=13} |
| 7 | casOut={name='call_features', caslib='casuser', replace=true}; |
| 8 | RUN; |
Expected Result
The output table 'call_features' should contain the 'call_id' column and the computed MFCC feature vectors for each audio frame. The number of cepstral coefficients per frame should be 13.