Personalize your experience with our AI based speaker ID. With an identification accuracy of more than 95% with very low CPU and memory requirement, its ideal for all personal devices such as phones, smart speakers, smart radios and laptops.
Speaker ID block diagram
Key KPIs
- TPR: 95% on android mobile phone for single user enrolment use cases.
- FPR: 5% on android mobile phone for single user enrolment.
- CPU: 370MHZ on Android Mobile Phone.
- Latency: less than 350 ms delay.
Competitive Benchmarking for Home use case
| Benchmarking | Enrolled/Impostor | True positive rate | False positive rate |
|---|---|---|---|
| Meeami Speaker ID (XMOS+P64) | (05)/(05) | 85% | 15% |
| Amazon Echo DOT | (05)/(05) | 98% | 30% |
Text Dependent Speaker ID
- Uses proprietary feature vector (40-D) and pre-trained background model @2.5 MB
- Operates on frame by frame basis
- Tunes the background model during enrollment
- Provision for updating the background model
- Identification and imposter rejection is 95% and above
- supports specific tuning for max identification or imposter rejection
- CPU : 360MHz (unoptimized), additional 1MHz per new speaker
Text Independent Speaker ID
- Operates once in 200 frames of speech
- Identification Accuracy (IA) - 90%
- Imposter Rejection Rate - 83%
- CPU: Base: 320 MHz (unoptimized)
- Model size : 38 MB
- Every increment in enrolled speaker consumes negligible MHz (around 0.5MHz)
- Average delay: ~2/3 seconds.
