Machine learning based upon feedback from contact center analysis

Abstract

A signal monitoring apparatus and method involving devices for monitoring signals representing communications traffic, devices for identifying at least one predetermined parameter by analyzing the context of the at least one monitoring signal, a device for recording the occurrence of the identified parameter, a device fur identifying the traffic stream associated with the identified parameter, a device for analyzing the recorded data relating to the occurrence, and a device, responsive to the analysis of the recorded data, for controlling the handling of communications traffic within the apparatus.

Description

BRIEF DESCRIPTION OF THE DRAWINGS(Step 302, FIG. 3). As will be appreciated by the arrows employed for the signal lines 24, 26, the high impedance tap 20 is arranged to monitor outgoing voice signals from the call-centre 10 whereas the high impedance tap 22 is arranged to monitor incoming signals to the call-centre 10. The voice traffic on the lines 24, 26 therefore form a two-way conversation between a call-centre operative using one of the terminals 12 and a customer (not illustrated).

The monitoring apparatus 16 embodying the present invention further includes a computer telephone link 28 whereby data traffic appearing at the exchange switch 14 can be monitored as required.

The digital voice recorder 18 is connected to a network connection 30 which can be in the form of a wide area network (WAN), a local area network (LAN) or an internal bus of a central processing unit of a computer.

Also connected to the network connection 30 is a replay station 32, a configuration management application station 34, a station 36 providing speech and/or data analysis engine(s) and also storage means comprising a first storage means 38 for the relevant analysis rules and the results obtained and a second storage means 40 for storage of the data and/or speech monitor.

FIG. 2 illustrates the typical format of a data packet 42 used in accordance with the present invention and which comprises a packet header 44 of typically 48 bytes and a packet header 46 of typically of 2000 bytes.

The packet header is formatted so as to include the packet identification 48, the data format 50, a date and time stamp 52, the relevant channel number within which the data arises 54, the gain applied to the signal 56 and the data length 58.

The speech, or other data captured in accordance with the apparatus of the present invention, is found within the packet body 46 and within the format specified within the packet header 44.

The high impedance taps 20, 22 offer little or no effect on the transmission lines 24, 26 and, if not in digital form, the monitored signal is converted into digital form. For example, when the monitored signal comprises a speech signal, the signal is typically convened to a pulse code modulated (PCM) signal or is compressed as an Adaptive Differential PCM (ADPCM) signal.

Further, where signals are transmitted at a constant rate, the time of the start of the recordings is identified, for example by voltage or activity detection, i.e. so-called “vox” level detection, and the time is recorded. With asynchronous data signals, the start time of a data burst, and optionally the intervals between characters, may be recorded in addition to the data characters themselves.

The purpose of this is to allow a computer system to model the original signal to appropriate values of time, frequency and amplitude so as to allow the subsequent identification of one or more of the various parameters arising in association with the signal (see, FIG. 4). The digital information describing the original signals is then analysed at station 36 (Step 304; FIG. 3), in real time or later, to determine the required set of metrics, i.e. parameters, appropriate to the particular application (Step 306; FIG. 3).

FIG. 3 is a flowchart of an example process 300 for monitoring communications traffic. At stage 302, signals representing communications traffic are monitored. For example, the digital voice recorder 18 can monitor two-way conversation traffic associated with the exchange switch 14. At stage 304, a predetermined parameter is identified by analyzing the content. For example, a digital signal processor programmed with an appropriate algorithm can identify the predetermined parameter. At stage 306, the occurrence of the identified parameter is recorded. For example, the first storage 38 (analysis rules and results) can store the occurrence of the identified parameter. At stage 308, the traffic stream associated with the parameter is identified. For example, the speech/data analysis engine 36 can identify the traffic stream. At stage 310, the recorded data relating to the occurrence is analyzed to highlight the identified parameter or to modify analysis algorithm used to identify the parameter and/or modify the analysis algorithm used to determine the parameters. For example, the speech/data analysis engine 36 can analyze the recorded data stored in the first storage 38 to highlight particularly good or bad interactions based on the identified parameter. Alternatively, feedback may be used to modify the algorithms used to perform analysis of the monitored signals.

A particular feature of the system is in recording the two directions of data transmission separately so allowing further analysis of information sent in each direction independently. In analogue telephone systems, this may be achieved by use of a four-wire (as opposed to two-wire) circuit whilst in digital systems, it is the norm to have the two directions of transmission separated onto separate wire pairs. In the data world, the source of each data packet is typically stored alongside the contents of the data packet.

A further feature of the system is in recording the level of amplification or attenuation applied to the original signal. This may vary during the monitoring of even a single interaction (e.g. through the use of Automatic Gain Control Circuitry). This allows the subsequent reconstruction and analysis of the original signal amplitude.

Another feature of the system is that monitored data may be “tagged” with additional information such as customer account numbers by an external system (e.g. the delivery of additional call information via a call logging port or computer telephony integration (CTI) port).

The importance of each of the parameters and the way in which they can be combined to highlight particularly good or bad interactions is defined by the user of the system (Step 308; FIG. 3). One or more such analysis profiles can be held in the system. These profiles determine the weighting given to each of the above parameters.

The profiles are normally used to rank a large number of monitored conversations and to identify trends, extremes, anomalies and norms. “Drill-down” techniques are used to permit the user to examine the individual call parameters that result in an aggregate or average score and, further, allow the user to select individual conversations to be replayed to confirm or reject the hypothesis presented by the automated analysis.

A particular variant that can be employed in any embodiment of the present invention uses feedback from the user's own scoring of the replayed calls to modify its own analysis algorithms (Step 310; FIG. 3). This may be achieved using neural network techniques or similar giving a system that learns from the user's own view of the quality of recordings.

A variant of the system uses its own and/or the scoring/ranking information to determine its further patterns of operation i.e.

• • determining which recorded calls to retain for future analysis,
  • determining which agents/lines to monitor and how often, and
  • determining which of the monitored signals to analyse and to what depth.

In many systems it is impractical to analyse all attributes of all calls hence a sampling algorithm may be defined to determine which calls will be analysed. Further, one or more of the parties can be identified (e.g. by calling-line identifier for the external party or by agent log-on identifiers for the internal party). This allows analysis of the call parameters over a number of calls handled by the same agent or coming from the same customer.

The system can use spare capacity on the digital signal processors (DSPs) that control the monitoring, compression or recording of the monitored signals to provide some or all of the analysis required. This allows analysis to proceed more rapidly during those periods when fewer calls are being monitored.

Spare CPU capacity on a PC at an agent's desk could be used to analyse the speech. This would comprise a secondary tap into the speech path being recorded as well as using “free” CPU cycles. Such an arrangement advantageously allows for the separation of the two parties, e.g. by tapping the headset/handset connection at the desk. This allows parameters relating to each party to be stored even if the main recording point can only see a mixed signal.

A further variant of the system is an implementation in which the systems recording and analysing the monitored signals are built into the system providing the transmission of the original signals (e.g. as an add-in card to an Automatic Call Distribution (ACD) system).

The apparatus illustrated is particularly useful for identifying the following parameters:

• • degree of interruption (i.e. overlap between agent talking and customer talking);
  • comments made during music or on-hold periods;
  • delays experienced by customers (i.e. the period from the end of their speech to an agent's response);
  • caller/agent talk ratios, i.e. which agents might be talking too much.

However, it should be appreciated that the invention could be adapted to identify parameters such as:

• • “relaxed/stressed” profile of a caller or agent (i.e. by determining changes in volume, speed and tone of speech)
  • frequency of keywords heard (separately from agents and from callers) e.g. are agents remembering to ask follow-up questions about a certain product/service etc; or how often do customers swear at each agent? Or bow often do agents swear at customers?
  • frequency of repeat calls. A combination of line, ID and caller ID can be provided to eliminate different people calling from single switchboard/business number
  • languages used by callers?
  • abnormal speech patterns of agents. For example if the speech recognition applied to an agent is consistently and unusually inaccurate for, say, half an hour, the agent should be checked for: drug abuse, excessive tiredness, drunkenness, stress, rush to get away etc.

It will be appreciated that the illustrated and indeed any embodiments of the present invention can be set up as follows.

The Digital Trunk Lines (e.g. T1/E1) can be monitored trunk side and the recorded speech tagged with the direction of speech. A MediaStar Voice Recorder chassis can be provided typically with one or two E1/T1 cards plus a number of DSP cards for the more intense speech processing requirements.

Much of its work can be done overnight and in time, some could be done by the DSPs in the mediastar's own cards: It is also necessary to remove or at least recognise, periods of music, on-hold periods, IVR rather than real agents speaking etc. thus, bundling with Computer Integrated Telephony Services such as Telephony Services API (TSAPI) in many cases is appropriate.

Analysis and parameter identification as described above can then be conducted. However, as noted, if it is not possible to analyse all speech initially, analysis of a recorded signal can be conducted.

In any case the monitoring apparatus may be arranged to only search initially for a few keywords although re-play can be conducted so as to look for other keywords.

It should be appreciated that the invention is not restricted to the details of the foregoing embodiment. For example, any appropriate form of telecommunications network, or signal transmission media, can be monitored by apparatus according to this invention and the particular parameters identified can be selected, and varied, as required.

Claims

1. A signal monitoring system for monitoring and analyzing communications passing through a monitoring point, the system comprising:

a digital voice recorder (18) for monitoring two-way conversation traffic streams passing through the monitoring point, said digital voice recorder having connections (20) for being operatively attached to the monitoring point;

a digital processor (30) connected to said digital voice recorder for identifying at least one predetermined parameter by analyzing the voice communication content of at least one monitored signal taken from the traffic streams;

a recorder (38) attached to said digital processor for recording occurrences of the predetermined parameter;

a traffic stream identifier (36) for identifying the traffic stream associated with the predetermined parameter;

a data analyzer (36) connected to said digital processor for analyzing the recorded data relating to the occurrences; and

a communication traffic controller (34) operatively connected to said data analyzer and, operating responsive to the analysis of the recorded data, for controlling the handling of communications traffic within said monitoring system.

2. The monitoring system of claim 1, wherein said at least one predetermined parameter includes a frequency of keywords identified in the voice communication content of the at least one monitored signal.

3. The monitoring system of claim 1, wherein said digital processor further identifies episodes of anger or shouting by analyzing amplitude envelope.

4. The monitoring system of claim 1, wherein said at least one predetermined parameter is a prosody of the voice communication content of the at least one monitored signal.

5. The monitoring system of claim 1, wherein said connections for being operatively attached to the telephony exchange switch are attached via high impedance taps (20) to telephone signal lines (24, 26) attached to said telephony exchange switch.

6. The monitoring system of claim 1, wherein said communication traffic controller serves to identify at least one section of traffic relative to another so as to identify a source of the predetermined parameter.

7. The monitoring system of claim 1, wherein said communication traffic controller serves to influence further monitoring actions within the apparatus.

8. The monitoring system of claim 1, wherein the analyzed contents of the at least one monitored signal comprise the interaction between at least two signals representing an at least two-way conversation.

9. The monitoring system of claim 1, wherein the recorder operates in real time to provide a real-time indication of the occurrence.

10. The monitoring system of claim 1, wherein said digital voice recorder comprises an analog/digital convertor (18) for converting analog voice into a digital signal.

11. The monitoring system of claim 1, wherein said digital processor is a Digital Signal Processor (30) arranged to operate in accordance with an analyzing algorithm.

12. The monitoring system of claim 1, wherein the digital processor is arranged to operate in real time.

13. The monitoring system of claim 1, further comprising a replay station (32) connected to said digital processor and arranged such that the voice communication content of the at least one monitored signal can be recorded and monitored by said digital processor for identifying the at least one parameter at some later time.

14. The monitoring system of claim 1, wherein the at least one predetermined parameter comprises plural predetermined parameters and wherein said recorder records the occurrence of the plural predetermined parameters in each of the two directions of traffic separately.

15. The monitoring system of claim 1, wherein said traffic stream identifier comprises a means for receiving an identifier tagged onto the traffic so as to identify its source.

16. The monitoring system of claim 1, wherein said digital voice recorder for monitoring the traffic streams is operative responsive to an output from said traffic stream identifier identifying the source of the conversation in which the predetermined parameter has been identified, or a threshold occurrence of the predetermined parameter has been exceeded.

17. The monitoring system of claim 1, wherein said digital voice recorder, said digital processor, said recorder, said traffic stream identifier, and said data analyzer reside on an add-in card to a telecommunications system.

18. A method for adaptively analyzing a call center interaction, comprising:

analyzing voice interaction data representative of the call center interaction at a speech/data analysis engine to determine a parameter associated with the voice interaction data; and

receiving scoring as feedback regarding the call center interaction to modify an analysis algorithm employed by the speech/data analysis engine, and from the feedback;

determining which recorded calls to retain for future analysis,

determining which agents/lines to monitor and how often, and

determining which monitored agents/lines to analyze and at least one decision rule associated with the monitoring.

19. The method of claim 18, wherein analyzing the voice interaction data comprises identifying voice communication content included in the voice interaction data.

20. The method of claim 19, wherein identifying voice communication content includes identifying a frequency of keywords identified in the voice interaction data.

21. The method of claim 19, wherein identifying voice communication content includes identifying episodes of anger or shouting based upon an amplitude envelope associated with the voice interaction data.

22. The method of claim 19, wherein identifying voice communication content includes identifying a prosody associated with the voice communication content of the voice interaction data.

23. The method of claim 19, wherein identifying voice communication content includes examining incoming and outgoing traffic streams to identify whether a talk-over condition exists with respect to the voice interaction data.

24. The method of claim 19, further comprising identifying whether one or more of a predetermined group of words exists with respect to the voice interaction data.

25. The method of claim 19, wherein identifying voice communication content includes identifying stress voice content associated with the voice interaction data.

26. The method of claim 25, wherein stress is identified by determining changes in volume, speed and tone of voice content associated with the voice interaction data.

27. The method of claim 19, wherein identifying voice communication content includes identifying a delay between voice transmissions in opposite directions.

28. The method of claim 19, further comprising scoring the telephone conversation at the replay station to analyze a customer experience during voice interactions.

29. The method of claim 18, further comprising defining said predetermined parameter as at least one of: a threshold frequency of at least one user defined keyword; a prosody associated with the voice interaction indicating stress and intonation in the voice interaction; or, anger evidenced by an amplitude envelope associated with the voice interaction.

30. An adaptive learning voice interaction recording and analysis system, comprising:

a recorder operable to acquire voice interaction data in a call center environment;

an analysis module operable to analyze voice communication content associated with the voice interaction data received from the recorder, the analysis module being operable to identify at least one predetermined parameter by analyzing voice communication content of at least one monitored signal taken from the voice interaction data, wherein the analysis module uses one or more analysis algorithms to analyze voice communication content;

a storage device operable to store analyzed voice communication content; and

a replay station operable to facilitate feedback by retrieving stored analyzed voice communication content from the storage device, wherein the feedback is used modify the one or more analysis algorithms.

31. The system of claim 30, wherein the analysis module automatically identifies an occurrence of said at least one predetermined parameter in the voice interaction data.

32. The system of claim 31, wherein analysis of the voice communication content comprises identifying voice communication content included in the voice interaction data.

33. The system of claim 32, wherein identifying voice communication content includes identifying a frequency of keywords identified in the voice interaction data.

34. The system of claim 32, wherein identifying voice communication content includes identifying episodes of anger or shouting based upon an amplitude envelope associated with the voice interaction data.

35. The system of claim 32, wherein identifying voice communication content includes identifying a prosody associated with the voice communication content of the voice interaction data.

36. The system of claim 32, wherein identifying voice communication content includes examining incoming and outgoing traffic streams to identify whether a talk-over condition exists with respect to the voice interaction data.

37. The system of claim 32, wherein identifying voice communication content includes identifying whether one or more of a predetermined group of words exists with respect to the voice interaction data.

38. The system of claim 32, wherein identifying voice communication content includes identifying voice stress associated with the voice interaction data.

39. The system of claim 38, wherein stress is identified by determining changes in volume, speed and tone of voice content associated with the voice interaction data.

40. The system of claim 32, wherein identifying voice communication content includes identifying a delay between voice transmissions in opposite directions.

41. The system of claim 30, wherein the analysis module is further operable to score the voice interaction data to analyze a customer experience during voice interactions.

42. The system of claim 41, wherein the analysis module scores the voice interaction data using such characteristics as delay between transmissions in opposite directions, anger, stress, or use of one or more of a group of words to analyze customer experience during the voice interaction.