with a Speech Recognizer

boorishadamantAI and Robotics

Oct 29, 2013 (3 years and 5 months ago)

50 views

Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

Evaluating Speech Separation
with a Speech Recognizer

Mikkel N. Schmidt

Technical University of Denmark

Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

The Speech Separation Problem


Separate a single
-
channel mixture of speech from
known speakers

Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

Evaluation of the Separation


Signal to Noise Ratio


Simple measure, has only indirect relation to percieved quality.


Representation
-
based Metrics


In systems based on time
-
frequency masking, evaluate the masks.


Perceptual Models


Promising.


High
-
level Attributes


For example word error rate in a speech recognition setup.


Listening
-
tests


Expensive and time
-
consuming.

Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

The GRID Corpus


Corpus size: 34 Speakers x 1000 Utterances


Example utterance: “Bin White By R 7 Please”



Test the separation, not the speech recognizer


Simple, structured sentences, small vocabulary


Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

The Speech Separation Challenge


Identify the
letter

and
number

in the sentence
spoken by the talker who says “White”

Informatics and Mathematical Modelling / Intelligent Signal Processing

Evaluating Speech Separation with a Speech Recognizer

Results


Mixtures at 0 dB. Speakers of opposite gender


Word error rate

No separation

68%

Our algorithm

29%

Human performance

14%


More advanced algorithms have been shown to
exceed human performance in certain conditions