IraCPA Machine Detection and Accuracy
Answering Machine Detection (AMD)
Answering Machine Detection (AMD) is a technology used with dialers to identify calls answered by voicemail systems instead of a live person. The objective is to quickly determine whether the call has been answered by:
- Live Human
- Answering Machine
- Voicemail
- Fax
- Modem
- IVR
This decision must typically be made within 2 seconds to comply with TCPA regulations in the United States. TCPA requires the call to be transferred to a live agent within 2 seconds after the called party completes their greeting.
Evaluation Techniques
Epicode's Call Progress Analysis (IraCPA) module uses advanced voice analysis algorithms that evaluate:
- Speech patterns
- Tone of voice
- Speaking speed
- Post-greeting silence
- Audio characteristics
These techniques enable industry-leading detection accuracy with an analysis time of only 1.75 seconds, leaving approximately 0.25 seconds for applications to transfer or bridge the call to an agent while remaining fully TCPA compliant.
Voicemail Beep Detection
IraCPA also supports voicemail drop workflows.
After detecting that the call has been answered by an answering machine:
- IraCPA immediately notifies the application.
- Audio analysis continues while listening for the voicemail beep.
- Once the beep is detected, another notification is sent.
- The application can then automatically play a prerecorded voicemail message.
Handling Early Answer Events
During large-scale deployments, some telecom carriers generate the call answered event before the audio path is fully established. This can result in the first few seconds of the call containing complete silence.
To improve detection accuracy, IraCPA can:
- Distinguish between absolute silence and ambient background noise
- Ignore initial silence for a configurable duration (up to 5 seconds)
- Begin analysis only after meaningful audio is detected
This significantly improves the accuracy of distinguishing between live humans and answering machines.
Benchmarks
The following benchmarks are based on extensive internal testing along with production deployments across the United States and Europe.
| Time Since First Utterance | Sample Size | Human Accuracy | Machine Accuracy | Fax Accuracy | Overall Accuracy |
|---|---|---|---|---|---|
| 1.50 seconds | 10,000 | 86% | 90% | 100% | 86.8% |
| 1.75 seconds | 10,000 | 92% | 95% | 100% | 93.2% |