TitaNet: Neural Model for speaker representation with 1D Depth-wise separable convolutions and global context
In this paper, we propose TitaNet, a novel neural network architecture for extracting speaker representations. We employ 1D depth-wise separable convolutions with Squeeze-and-Excitation (SE) layers with global context followed by channel attention based statistics pooling layer to map variable-length utterances to a fixed-length embedding (t-vector). TitaNet is a scalable architecture and achieves state-of-the-art performance on speaker verification task with an equal error rate (EER) of 0.68% on the VoxCeleb1 trial file and also on speaker diarization tasks with diarization error rate (DER) of 1.73% on AMI-MixHeadset, 1.99% on AMI-Lapel and 1.11% on CH109. Furthermore, we investigate various sizes of TitaNet and present a light TitaNet-S model with only 6M parameters that achieve near state-of-the-art results in diarization tasks.
PDF AbstractCode
Colab
LibriSpeech
VoxCeleb1
VoxCeleb2
| Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Uses Extra Training Data |
Benchmark |
|---|---|---|---|---|---|---|---|
| Speaker Diarization | AMI Lapel | TitaNet-L (NME-SC) | DER(%) | 2.03 | # 3 | ||
| Speaker Diarization | AMI Lapel | ECAPA (SC) | DER(%) | 2.36 | # 4 | ||
| Speaker Diarization | AMI Lapel | TitaNet-S (NME-SC) | DER(%) | 2.00 | # 2 | ||
| Speaker Diarization | AMI Lapel | TitaNet-M (NME-SC) | DER(%) | 1.99 | # 1 | ||
| Speaker Diarization | AMI MixHeadset | TitaNet-M (NME-SC) | DER(%) | 1.79 | # 3 | ||
| Speaker Diarization | AMI MixHeadset | TitaNet-S (NME-SC) | DER(%) | 2.22 | # 4 | ||
| Speaker Diarization | AMI MixHeadset | ECAPA (SC) | DER(%) | 1.78 | # 2 | ||
| Speaker Diarization | AMI MixHeadset | TitaNet-L (NME-SC) | DER(%) | 1.73 | # 1 | ||
| Speaker Diarization | CALLHOME-109 | titanet-s | DER(%) | 1.11 | # 1 | ||
| Speaker Diarization | CH109 | TitaNet-M (NME-SC) | DER(%) | 1.13 | # 2 | ||
| Speaker Diarization | CH109 | TitaNet-L (NME-SC) | DER(%) | 1.19 | # 3 | ||
| Speaker Diarization | CH109 | x-vector (PLDA + AHC) | DER(%) | 9.72 | # 4 | ||
| Speaker Diarization | CH109 | TitaNet-S (NME-SC) | DER(%) | 1.11 | # 1 | ||
| Speaker Diarization | NIST-SRE 2000 | x-vector (PLDA + AHC) | DER(%) | 8.39 | # 5 | ||
| Speaker Diarization | NIST-SRE 2000 | TitaNet-S (NME-SC) | DER(%) | 6.37 | # 2 | ||
| Speaker Diarization | NIST-SRE 2000 | TitaNet-L (NME-SC) | DER(%) | 6.73 | # 4 | ||
| Speaker Diarization | NIST-SRE 2000 | x-vector (MCGAN) | DER(%) | 5.73 | # 1 | ||
| Speaker Diarization | NIST-SRE 2000 | TitaNet-M (NME-SC) | DER(%) | 6.47 | # 3 | ||
| Speaker Verification | VoxCeleb | TitanNet -S | EER | 1.15 | # 19 | ||
| Speaker Verification | VoxCeleb | TitanNet -L | EER | 0.68 | # 13 | ||
| Speaker Verification | VoxCeleb | TitanNet -M | EER | 0.81 | # 15 |
