Academia.eduAcademia.edu

Figure 4 – uploaded by Jindřich Žďánský

See full PDFdownloadDownload figure
Fig. 3. Accuracy in the task of the SOI dominance identification; each marke corresponds to a different SNR.

Figure 3 Accuracy in the task of the SOI dominance identification; each marke corresponds to a different SNR.

Related Figures (15)

The update rules for finding the optimum point of the auxiliary contrast function are obtained in the form:
The update rules for finding the optimum point of the auxiliary contrast function are obtained in the form:
The input sizes for the context layers are stated after the mean pooling operation. DESCRIPTION OF THE FSMN PRODUCING THE X- VECTORS
The input sizes for the context layers are stated after the mean pooling operation. DESCRIPTION OF THE FSMN PRODUCING THE X- VECTORS
ig. 1. | Source trajectories and locations for the Dynamic dataset. square subtraction of the extracted signal a = (wht)F tee from x. Let ar ” be the mixing vector after 2 deflation steps computed mm the tth block via (8). Due to the orthogonality of w*:? and a® * the subtraction is achieved through
ig. 1. | Source trajectories and locations for the Dynamic dataset. square subtraction of the extracted signal a = (wht)F tee from x. Let ar ” be the mixing vector after 2 deflation steps computed mm the tth block via (8). Due to the orthogonality of w*:? and a® * the subtraction is achieved through
present. The prolongation of inner blocks does not bring the ad- vantage of more available data. Application of an insufficiently short block (50 frames) allows for good adaptation ow SOI Attenuation), but the overall IS suppression is deteriorating (low iSIR).
present. The prolongation of inner blocks does not bring the ad- vantage of more available data. Application of an insufficiently short block (50 frames) allows for good adaptation ow SOI Attenuation), but the overall IS suppression is deteriorating (low iSIR).
DYNAMIC DATASET: THE EXTRACTION PERFORMANCE OF THE CSV-AUXIVE (CSV), THE STATIC (FS-IVE) AND THE BLOCK-WISE STATIC (BS-IVE) IVE TECHNIQUES Fig. 2. Accuracy in the task of the dominant speaker identification; each marker corresponds to a different SNR.
DYNAMIC DATASET: THE EXTRACTION PERFORMANCE OF THE CSV-AUXIVE (CSV), THE STATIC (FS-IVE) AND THE BLOCK-WISE STATIC (BS-IVE) IVE TECHNIQUES Fig. 2. Accuracy in the task of the dominant speaker identification; each marker corresponds to a different SNR.
The subscript Ly denotes the number of frames within the analyzed block.
The subscript Ly denotes the number of frames within the analyzed block.
the Norwegian dataset. This decrease does not stem from the language as such but it is caused by longer silences between Norwegian sentences. When the SOI is quiet, the non-piloted extractor tends to converge to an arbitrary active source. The utilization of a pilot completely removes this difference. The results for CSV-AuxIVE piloted via g*Y®° are comparable for both datasets (difference is maximally 1 dB in iSIR and iSDR); the proposed method can thus be considered language independent in this experiment.
the Norwegian dataset. This decrease does not stem from the language as such but it is caused by longer silences between Norwegian sentences. When the SOI is quiet, the non-piloted extractor tends to converge to an arbitrary active source. The utilization of a pilot completely removes this difference. The results for CSV-AuxIVE piloted via g*Y®° are comparable for both datasets (difference is maximally 1 dB in iSIR and iSDR); the proposed method can thus be considered language independent in this experiment.
DYNAMIC DATASET: THE EXTRACTION PERFORMANCE OF PILOTED CSV-AUXIVE (Lp = 200) WITH RESPECT TO X-VECTOR CONTEXT Le Fig. 4. Accuracy in the task of the SOI dominance identification with respect to language of speakers in the enrollment set; LZ. = 11 and each marker corresponds to a different SNR.
DYNAMIC DATASET: THE EXTRACTION PERFORMANCE OF PILOTED CSV-AUXIVE (Lp = 200) WITH RESPECT TO X-VECTOR CONTEXT Le Fig. 4. Accuracy in the task of the SOI dominance identification with respect to language of speakers in the enrollment set; LZ. = 11 and each marker corresponds to a different SNR.
Fig. 5. MC-WSJ0-2mix, 4 channels: iSDR distributions achieved by CSV- AuxIVE endowed with various forms of guidance.
Fig. 5. MC-WSJ0-2mix, 4 channels: iSDR distributions achieved by CSV- AuxIVE endowed with various forms of guidance.
Fig.6. _MC-WSJ0-2mix, 4 channels: SDR achieved by the CSV-AuxIVE using oracle pilot, whose frames corresponding to SOJ are gradually interchanged with frames dominated by the interfering source.
Fig.6. _MC-WSJ0-2mix, 4 channels: SDR achieved by the CSV-AuxIVE using oracle pilot, whose frames corresponding to SOJ are gradually interchanged with frames dominated by the interfering source.
Significant cases denote samples in which the erroneous increase/decrease in SIR/SDR is larger than 1 dB, 0.01 in STOI or 0.05 in PESQ. MC-WSJ0-2MIx (SPEECH-SPEECH MIXTURES): THE EVALUATION OF THE PROPOSED EXTRACTION ASSESSMENT SERVING AS A BINARY CLASSIFIER OF THE SPEECH-QUALITY METRICS
Significant cases denote samples in which the erroneous increase/decrease in SIR/SDR is larger than 1 dB, 0.01 in STOI or 0.05 in PESQ. MC-WSJ0-2MIx (SPEECH-SPEECH MIXTURES): THE EVALUATION OF THE PROPOSED EXTRACTION ASSESSMENT SERVING AS A BINARY CLASSIFIER OF THE SPEECH-QUALITY METRICS
Significant cases denote samples in which the erroneous increase/decrease in SIR/SDR is larger than 1 dB, 0.01 in STOI or 0.05 in PESQ. CHIME-4 (SIMULATED DEVELOPMENT PART, SPEECH-NOISE MIXTURES): THE EVALUATION OF THE PROPOSED EXTRACTION ASSESSMENT SERVING AS A BINARY CLASSIFIER OF THE SPEECH-QUALITY METRICS
Significant cases denote samples in which the erroneous increase/decrease in SIR/SDR is larger than 1 dB, 0.01 in STOI or 0.05 in PESQ. CHIME-4 (SIMULATED DEVELOPMENT PART, SPEECH-NOISE MIXTURES): THE EVALUATION OF THE PROPOSED EXTRACTION ASSESSMENT SERVING AS A BINARY CLASSIFIER OF THE SPEECH-QUALITY METRICS
Fig. 8. |. MC-WSJO-2mix: dependency between the improvements of the objec- tive criteria and the improvements of PLDA score on speech-speech mixtures.
Fig. 8. |. MC-WSJO-2mix: dependency between the improvements of the objec- tive criteria and the improvements of PLDA score on speech-speech mixtures.
Fig. 7. CHiME-4 (simulated development part): dependency between the improvements of the objective criteria and the improvements of PLDA score on speech-noise mixtures.
Fig. 7. CHiME-4 (simulated development part): dependency between the improvements of the objective criteria and the improvements of PLDA score on speech-noise mixtures.
The column “Tr. data” quantifies the volume of the required scenario-specific training data. MC-WSJ0-2m1x: SDR [DB] YIELDED USING MACHINE-LEARNING (ML), BLIND SOURCE SEPARATION (BSS) AND BLIND SOURCE EXTRACTION (BSE) TABLE IX
The column “Tr. data” quantifies the volume of the required scenario-specific training data. MC-WSJ0-2m1x: SDR [DB] YIELDED USING MACHINE-LEARNING (ML), BLIND SOURCE SEPARATION (BSS) AND BLIND SOURCE EXTRACTION (BSE) TABLE IX
Related topics:
Computer ScienceArtificial Intelligence
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved