Monaural Speech Segregation: Representation, Pitch, and Amplitude Modulation DeLiang Wang The Ohio State University.

Monaural Speech Segregation: Representation, Pitch,

and Amplitude Modulation

DeLiang Wang

The Ohio State University

Outline of Presentation

Introduction Speech segregation problem Auditory scene analysis (ASA) approach

A multistage model for computational ASA On amplitude modulation and pitch tracking Oscillatory correlation theory for ASA

Speech Segregation Problem

In a natural environment, target speech is usually corrupted by acoustic interference. An effective system for speech segregation has many applications, such as automatic speech recognition, audio retrieval, and hearing aid design

Most speech separation techniques require multiple sensors

Speech enhancement developed for the monaural situation can deal with only specific acoustic interference

Auditory Scene Analysis (Bregman’90)

Listeners are able to parse the complex mixture of sounds arriving at the ears in order to retrieve a mental representation of each sound source

ASA would take place in two conceptual processes:

Segmentation. Decompose the acoustic mixture into sensory elements (segments)

Grouping. Combine segments into groups, so that segments in the same group are likely to have originated from the same environmental source

Auditory Scene Analysis - continued

The grouping process involves two aspects: Primitive grouping. Innate data-driven mechanisms,

consistent with those described by Gestalt psychologists for visual perception (proximity, similarity, common fate, good continuation, etc.)

Schema-driven grouping. Application of learned knowledge about speech, music and other environmental sounds

Computational Auditory Scene Analysis

Computational ASA (CASA) systems approach sound separation based on ASA principles

Weintraub’85, Cooke’93, Brown & Cooke’94, Ellis’96, Wang’96

Previous CASA work suggests that: Representation of the auditory scene is a key issue Temporal continuity is important (although it is ignored

in most frame-based sound processing algorithms) Fundamental frequency (F0) is a strong cue for

grouping

A Multi-stage Model (Wang & Brown’99)

NeuralOscillatorNetwork

Correlogram

Cross-channelCorrelation

ResynthesisHairCells

CochlearFiltering

Speechand Noise

ResynthesizedSpeech

ResynthesizedNoise

Correlogram (detail)

TimeLag

Neural Oscillator Network (detail)

GroupingLayer

SegmentationLayer

GlobalInhibitor

TimeFrequency

Auditory Periphery Model

A bank of fourth-order gammatone filters (Patterson et al.’88)

Meddis hair cell model converts gammatone output to neural firing

Auditory Periphery - Example

Hair cell response to utterance: “Why were you all weary?” mixed with phone ringing

128 filter channels arranged in ERB

Time (seconds)0.0 1.5

5000

2741

1457

729

315

80

Mid-level representations form the basis for segment formation and subsequent grouping

Correlogram extracts periodicity information from simulated auditory nerve firing patterns

Summary correlogram is used to identify F0 Cross-correlation between adjacent

correlogram channels identifies regions that are excited by the same frequency component or formant

Mid-level Auditory Representations

Mid-level Representations - Example

Correlogram and cross-channel correlation for the speech/telephone mixture

Oscillator Network: Segmentation Layer

Horizontal weights are unity, reflecting temporal continuity, and vertical weights are unity if cross-channel correlation exceeds a threshold, otherwise 0

A global inhibitor ensures that different segments have different phases

A segment thus formed corresponds to acoustic energy in a local time-frequency region that is treated as an atomic component of an auditory scene

Segmentation Layer - Example

Time (seconds)0.0 1.5

5000

2741

1457

729

315

80

Output of the segmentation layer in response to the speech/telephone mixture

Oscillator Network: Grouping Layer

At each time frame, an F0 estimate from the summary correlogram is used to classify channels into two categories; those that are consistent with the F0, and those that are not

Connections are formed between pairs of channels: mutual excitation if the channels belong to the same F0 category, otherwise mutual inhibition

Strong excitation within each segment The second layer embodies the grouping stage of

ASA

Grouping Layer - Example

Two streams emerge from the grouping layer at different times or with different phases Left: Foreground (original mixture ) Right: Background

Time (seconds)0.0 1.5

5000

2741

1457

729

315

80

Time (seconds)0.0 1.5

5000

2741

1457

729

315

80

Previous systems, including the Wang-Brown model, have difficulty in

Dealing with broadband high-frequency mixtures Performing reliable pitch tracking for noisy speech Retaining high-frequency energy of the target speaker

Our next step considers perceptual resolvability of various harmonics

Challenges Facing CASA

Resolved and Unresolved Harmonics

For voiced speech, lower harmonics are resolved while higher harmonics are not

For unresolved harmonics, the envelopes of filter responses fluctuate at the fundamental frequency of speech

Hence we apply different grouping mechanisms for low-frequency and high-frequency signals:

Low-frequency signals are grouped based on periodicity and temporal continuity

High-frequency signals are grouped based on amplitude modulation (AM) and temporal continuity

Proposed System (Hu & Wang'02)

Unit Labeling

Mixture

Peripheral and

mid-level processing

Initial Segregation

Resynthesis

Segregated speech

Pitch Tracking

Final Segregation

Envelope Representations - Example

(a) Correlogram and cross-channel correlation of hair cell response to clean speech

(b) Corresponding representations for response envelopes

Initial Segregation

The Wang-Brown model is used in this stage to generate segments and select the target speech stream

Segments generated in this stage tend to reflect resolved harmonics, but not unresolved ones

Pitch Tracking

Pitch periods of target speech are estimated from the segregated speech stream

Estimated pitch periods are checked and re-estimated using two psychoacoustically motivated constraints:

Target pitch should agree with the periodicity of the time-frequency (T-F) units in the initial speech stream

Pitch periods change smoothly, thus allowing for verification and interpolation

Pitch Tracking - Example

(a) Global pitch (Line: pitch track of clean speech) for a mixture of target speech and ‘cocktail-party’ intrusion

(b) Estimated target pitch

T-F Unit Labeling

In the low-frequency range: A T-F unit is labeled by comparing the periodicity of its

autocorrelation with the estimated target pitch

In the high-frequency range: Due to their wide bandwidths, high-frequency filters

generally respond to multiple harmonics. These responses are amplitude modulated due to beats and combinational tones (Helmholtz, 1863)

A T-F unit in the high-frequency range is labeled by comparing its AM repetition rate with the estimated target pitch

AM - Example

(a) The output of a gammatone filter (center frequency: 2.6 kHz) to clean speech

(b) The corresponding autocorrelation function

AM Repetition Rates

To obtain AM repetition rates, a filter response is half-wave rectified and bandpass filtered

The resulting signal within a T-F unit is modeled by a single sinusoid using the gradient descent method. The frequency of the sinusoid indicates the AM repetition rate of the corresponding response

Final Segregation

New segments corresponding to unresolved harmonics are formed based on temporal continuity and cross-channel correlation of response envelopes (i.e. common AM). Then they are grouped into the foreground stream according to AM repetition rates

The foreground stream is adjusted to remove the segments that do not agree with the estimated target pitch

Other units are grouped according to temporal and spectral continuity

Ideal Binary Mask for Performance Evaluation

Within a T-F unit, the ideal binary mask is 1 if target energy is stronger than interference energy, and 0 otherwise

Motivation: Auditory masking - stronger signal masks weaker one within a critical band

Further motivation: Ideal binary masks give excellent listening experience and automatic speech recognition performance

Thus, we suggest to use ideal binary masks as ground truth for CASA performance evaluation

Monaural Speech Segregation Example

Left: Segregated speech stream (original mixture: ) Right: Ideal binary mask

Systematic Evaluation

Evaluated on a corpus of 100 mixtures (Cooke’93): 10 voiced utterances x 10 noise intrusions

Noise intrusions have a large variety

Resynthesis stage allows estimation of target speech waveform

Evaluation is based on ideal binary masks

Signal-to-Noise Ratio (SNR) Results

Average SNR gain: 12.1 dB; average improvement over Wang-Brown: 5 dB Major improvement occurs in target energy retention, particularly in the high-

frequency range

Segregation Examples

Mixture

Ideal Binary Mask

Wang-Brown

New System

How Does Auditory System Perform ASA?

Information about acoustic features (pitch, spectral shape, interaural differences, AM, FM) is extracted in distributed areas of the auditory system

Binding problem: How are these features combined to form a perceptual whole (stream)?

Hierarchies of feature-detecting cells exist, but do not seem to constitute a solution to the binding problem

Oscillatory Correlation Theory (von der Malsburg & Schneider’86; Wang’96)

Neural oscillators are used to represent auditory features

Oscillators representing features of the same source are synchronized (phase-locked with zero phase lag), and are desynchronized from oscillators representing different sources

Supported by growing experimental evidence, e.g. oscillations in auditory cortex measured by EEG, MEG and local field potentials

Oscillatory Correlation Representation

FD: Feature

Detector

FD1FDnFD2FDaFDbFDxTimeTime

Oscillatory Correlation for ASA

LEGION dynamics (Terman & Wang’95) provides a computational foundation for the oscillatory correlation theory

The utility of oscillatory correlation has been demonstrated for speech separation (Wang-Brown’99), modeling auditory attention (Wrigley-Brown’01), etc.

Issues

Grouping is entirely pitch-based, hence limited to segregating voiced speech

How to group unvoiced speech?

Target pitch tracking in the presence of multiple voiced sources

Role of segmentation We found increased robustness with segments as an

intermediate representation between streams and T-F units

Summary

Multistage ASA approach to monaural speech segregation

Performs substantially better than previous CASA systems

Oscillatory correlation theory for ASA Key issue is integration of various grouping cues

Collaborators

Recent work with Guoning Hu- Ohio State University

Earlier work with Guy Brown - University of Sheffield