Microphone Array Processing : A Quick Update Iain McCowan - - PowerPoint PPT Presentation

Microphone Array Processing : A Quick Update

Iain McCowan Guillaume Lathoud, Darren Moore, Olivier Masson

TAM May 2003 – p. 1/6

Outline

• Speech enhancement
• Speaker segmentation
• Files available online

TAM May 2003 – p. 2/6

Speech enhancement

• Improving enhancement in overlapping speech by post-filtering

beamformer outputs

• Beamformer outputs : yn(f) for each speaker location n = 1 : N
• Post-filter A (Wiener-like - |S|2

|N|2 )

• ˆ

yn(f) = |yn(f)|2

1 N−1

• m=n |ym(f)|2 yn(f)

(1)

• Post-filter B (Binary Mask)
• ˆ

yn(f) =

• yn(f)

n = arg maxm ym(f)

• therwise

(2)

TAM May 2003 – p. 3/6

Speech enhancement

• Subjectively, post-filter B leads to significant reduction in

cross-talk level.

• To verify, initial recognition experiments
• MONC (Multi-channel Overlapping Numbers Corpus -

re-recording of Numbers 95). Note : baseline lapel with no conflicting

speech is 7.0% WER.

• With one overlapping speaker (word error rates) :

Lapel Previous Array Best Post-filter B 26.7 19.3 12.2

• With two overlapping speaker :

Lapel Previous Array Best Post-filter B 35.3 26.6 15.8

TAM May 2003 – p. 4/6

Speaker Segmentation

• Previously, presented work on segmenting using location

features.

• Since then...
• Now doing clustering and segmentation using both location

features and standard acoustic features across meetings.

• Segment in terms of location and identity (cluster index)

concurrently.

• Using multi-stream HMM to cluster in each space

independently, but enforce same temporal segmentation.

• Automatically converges to correct number of locations and

identities.

• Initial results show high segmentation accuracy (≈ 95%

frame accuracy).

TAM May 2003 – p. 5/6

Files available online

• Now appearing on mmm.idiap.ch
• Beamformer outputs for Post-filter A and B for each seated

speaker location (1-4). (Scripted Meeting set only).

• Beamformer-B files have lower noise, though perhaps more

distortion than Beamformer-A.

• Beamformer outputs for whiteboard and presentation not yet

available.

• current beamformers are too precise for the typical movement in these

regions - investigating minimum beam-width constraint or adaptive techniques.

• Beamformer-B mix file available (BeamB-mix) - simple sum
• f 4 speaker beamformers.
• remember, this does not yet cater for white-board or presentation speech.
• currently, low level buzz apparent in this mix file... to be fixed.

TAM May 2003 – p. 6/6