Near-end Listening Enhancement Algorithms

Transcription

1 Near-end Listening Enhancement Algorithms Approaches for measurement and evaluation Jan Reimes HEAD acoustics GmbH Vienna, 2015/10/21

2 Overview Introduction Detection & Measurement Recording Procedure Measurement Example Analysis Proposals Speech Intelligibility vs. Quality Trade-Off Issue Proposal for Auditory Evaluation Conclusion 2

3 Introduction (1/2) Mobile phone calls in noisy environments may be extremely stressful for the person located at the near-end side. An improvement for this issue is proposed by recently developed so-called near-end listening enhancement (NELE) algorithms. These algorithms perform different manipulations on the down-link speech signal to decrease listening effort and improve intelligibility for the listener. 3

4 Introduction (2/2) Requirements for measurement (trends in standardization) Usage of real speech as test signal Realistic background noise scenarios / playback should be used for measurement Black-Box-Approach : No internals of a device-under-test are available, non-intrusive measurements only. Existing metrics which cannot be used for evaluation: STIPA, STITEL, RASTI, etc. (artificial test signals) Speech intelligibility index (needs processed but noise-free signal) ITU-T Rec. P.862 / P.863 (noise / near-end noise explicitly excluded in scope) ETSI EG / TS (specified only for sending direct and noise reduction scenarios) Neither technical nor psycho-acoustical metrics are available. 4

5 Recording Procedure (Down-Link, RCV) 1. Speech only (S) 2. BGN only (N) 3. BGN + Speech (SN) 4. Mix Signals (M = S + N, offline) NELE inactive: NELE active: Quantify differences SN vs. M 5

6 Measurement Example Mix: Device A: Recording: Mix: Device B: Recording: 6

7 Analysis Proposals Time Signals Delta-Level Delta-Spectrum Device B Device A 7

8 Speech Intelligibility vs. Quality (1/3) Some trade-offs must be regarded: Level increase may violate loudness rating requirement! Tuning towards intelligibility may decrease speech quality. Intelligibility may be more important in noisy scenario (ecall!). Proposal for auditory evaluation to quantify trade-off: Mixture of ITU-T P.800 & P.835 test with two scales: Speech quality (SQ, classic metric) Listening Effort (LE, substitute for intelligibility) Intelligibility vs. listening effort [Reimes et al., DAGA 2014]: Correlation is given under some circumstances Assessment of listening effort is easier & reproducible Assessment of intelligibility may depend on test material (?) 8

9 Speech Intelligibility vs. Quality (2/3) Some results of [Reimes et al., DAGA2014]: Several acoustic scenarios (handset, hands-free, ICC) 18 different conditions (bandwidth, listener position) Several car noise and silence conditions No NELE processings included, but everyday life situations Intelligibility test first, then combined LE/SQ category test. Intelligibility vs. listening effort (polynomial mapping): 9

10 Speech Intelligibility vs. Quality (3/3) Some results of [Reimes et al., DAGA2014]: Speech quality vs. listening effort: Bad LE & good SQ Large range of LE possible for constant SQ Large range of SQ possible for constant LE Dimensions are not correlated Borderline cases: data cannot be expected here Bad SQ & good LE 10

11 Conclusions Summary: Overview about NELE algorithms was given. Need for objectifying / quantifying performance of devices. Proposal for measurement setup and first analyses were given. Trade-off between speech quality and intelligibility. Auditory evaluations may be neccessary, proposal was presented. Outlook: Extension to hands-free / loudspeaker / public address system scenarios: binaural recording (and analysis)? Impact of NELE on different bandwidths (NB, WB, SWB, FB) Instrumental method for estimating combined listening effort and speech quality. 11

12 Jan Reimes Development Telecom Copyright HEAD acoustics GmbH