Note: Descriptions are shown in the official language in which they were submitted.
CA 02520594 2006-09-06
24080-748D
Voice Transcription System With Voice Playback and Silence
Compression
This application is a divisional of Canadian
Patent Application Serial No. 2,372,061 filed February 15,
2002.
BACKGROUND OF THE INVENTION
[002] The present invention relates to systems for
transcribing voice communications into text and specifically
to a system facilitating real-time editing of a transcribed
text stream by a human call assistant for higher accuracy.
[003] A system for real-time transcription of remotely
spoken voice signals is described in U.S. Patent 5,909,482
assigned to the same assignee as the present invention.
This system may find use implementing both a "captel"
(caption telephone) in which a user receives both voice and
transcribed text through a "relay" from a remote second
party to a conversation, and a "personal interpreter" in
which a user receives, through the relay, a text
transcription of words originating from the location of the
user.
[004] In either case, a human "call assistant" at the
relay listens to the voice signal and "revoices" the words
to a speech recognition computer program tuned to that call
assistant's voice. Revoicing is an operation in which the
call assistant repeats, in slightly delayed fashion, the
words she or he hears. The text output by the speech
recognition system is then transmitted to the captel or
personal interpreter. Revoicing by the call assistant
overcomes a current limitation of computer speech
recognition programs that they currently need to be trained
1
CA 02520594 2006-09-06
24080-748D
to a particular speaker and thus, cannot currently handle
direct translation of speech from a variety of users.
[005] Even with revoicing and a trained call assistant,
some transcription errors may occur, and therefore, the
above-referenced patent also discloses an editing system in
which the transcribed text is displayed on a computer screen
for review by the call assistant.
BRIEF SUMMARY OF THE INVENTION
[006] The present invention provides for a number of
improvements in the editing system described in the above-
referenced patent to speed and simplify the editing process
and thus generally improve the speed and accuracy of the
transcription. Most generally, the invention allows the
call assistant to select those words for editing based on
their screen location, most simply by touching the word on
the screen. Lines of text are preserved intact as they
scroll off the screen to assist in tracking individual words
and words on the screen change color to indicate their
status for editing and transmission. The delay before
transmission of transcribed text may be adjusted, for
example, dynamically based on error rates, perceptual rules,
or call assistant or user preference.
[007] The invention may be used with voice carryover in
a caption telephone application or for a personal
interpreter or for a variety of transcription purposes. As
described in the parent application, the transcribed voice
signal may be buffered to allow the call assistant to
accommodate varying transcription rates, however, the
present invention also provides more sophisticated control
of this buffering by the call assistant, for example adding
a foot control pedal, a graphic buffer gauge and automatic
2
CA 02520594 2006-09-06
24080-748D
buffering with invocation of the editing process. Further,
the buffered voice signal may be processed for "silence
compression" removing periods of silence. How aggressively
silence is removed may be made a function of the amount of
signal buffered.
[008] The invention further contemplates the use of
keyboard or screen entry of certain standard text in
conjunction with revoicing particularly for initial words of
a sentence which tend to repeat.
[009] The above aspects of the inventions are not
intended to define the scope of the invention for which
purpose claims are provided. Not all embodiments of the
invention will include all of these features.
A broad aspect of the invention provides a voice
transcription system comprising: an input circuit receiving
a voice signal including at least one spoken word from a
remote source and providing the spoken word to a call
assistant; a keyboard operable by a call assistant providing
for the input of text by the call assistant corresponding to
the voice signal, the input text including a text word
corresponding to the spoken word; output circuit
transmitting the input text to a remote user; wherein the
input circuit further includes a recording device recording
the voice signal and having playback of the voice signal
controllable by the call assistant to buffer the voice
signal; and a display device providing an indication of the
amount of voice signal buffered by the recording device
between a recording and playback point.
[010] In the following description, reference is made to
the accompanying drawings, which form a part hereof, and in
2a
CA 02520594 2006-09-06
24080-748D
which there is shown by way of illustration, a preferred
embodiment of the invention. Such embodiment
2b
CA 02520594 2002-02-15
24080-748D
also does not define the scope of the invention and reference must be made
therefore to the claims for this purpose.
BRIEF DESCRIPTION OF THE DRAWINGS
[011 ] Fig. 1 is a schematic diagram of a voice relay used with a captioned
telephone such as may make use of the present invention and showing a call
assistant receiving a voice signal for revoicing to a computer speech
recognition
program and reviewing the transcribed text on a display terminal;
[012] Fig. 2 is a figure similar to that of Fig. 1 showing a relay used to
implement a personal interpreter in which the speech signal and the return
text are
received and transmitted to a single location;
[0131 Fig. 3 is a simplified elevational view of the terminal of Figs. 1 and
2 as viewed by the call assistant;
[014] Fig. 4 is a generalized block diagram of the computer system of
Figs. 1 and 2 used for one possible implementation of the present invention
according to a stored program;
10151 Fig. 5 is a pictorial representation of a buffer system receiving a
voice signal prior to transcription by the call assistant such as may be
implemented by the computer of Fig. 4;
10161 Fig. 6 is a flowchart showing the elements of the program of Fig. 4
such as may realize the present invention including controlling the aging of
transcribed text prior to transmission;
10171 Fig. 7 is a detailed view of one flowchart block of Fig. 6 such as
controls the aging of text showing various inputs that may affect the aging
time;
[018] Fig. 8 is a graphical representation of the memory of the computer
of Fig. 4 showing data structures and programs used in the implementation of
the
present invention; and
[019] Fig. 9 is a fragmentary view of a caption telephone of Fig. I
showing a possible implementation of a user control for controlling a
transcription
speed accuracy tradeoff.
DETAILED DESCRIPTION OF THE INVENTION
CA 02520594 2002-02-15
24080-748D
[020] Referring now to Fig. 1, a relay 10, pennitting a hearing user 12 to
converse with a deaf or hearing impaired user 14, receives a voice signal 16
from
the mouthpiece of handset 13 of the hearing user 12. The voice signa116 is
processed by the relay 10 to produce a text stream signal 20 sent to the deaf
or
hearing impaired user 14 where it is displayed at a user terminal 22.
Optionally, a
modified voice signal 24 may also be provided to the earpiece of a handset 26
used by the deaf or hearing impaired user 14.
[0211 The deaf or hearing impaired user 14 may reply via a keyboard 28
per conventional relay operation through a connection (not shown for clarity)
or
may reply by spoken word into the mouthpiece of handset 26 to produce voice
signal 30. The voice signal 30 is transmitted directly to the earpiece of
handset 13
of the hearing user 12.
[0221 The various signals 24, 20 and 30 may travel through a single
conductor 32 (by frequency division multiplexing or data multiplexing
techniques
known in the art) or may be separate conductors. Equally, the voice signa130
and
voice signal 16 may be a single telephone line 34 but may also be multiple
lines.
[023] In operation, the relay 10 receives the voice signal 16 at computer
18 through an automatic gain control 36 providing an adjustment in gain to
compensate for various attenuations of the voice signal 16 in its
transmission. It is
then combined with an attenuated version of the voice signa130 (the other half
of
the conversation) arriving via attenuator 23. The voice signal 30 provides the
call
assistant 40 with context for a transcribed portion of the conversation. The
attenuator 23 modifies the voice signal 30 so as to allow the call assistant
40 to
clearly distinguish it from the principal transcribed conversation from user
12.
Other forms of discriminating between these two voices may be provided
including, for example, slight pitch shifting or filtering.
[024] The combined voice signals 16 and 30 are then received by a
"digital tape recorder" 19 and output after buffering by the recorder 19 as
headphone signal 17 to the earpiece of a headset 38 worn by a call assistant
40.
The recorder 19 can be controlled by a foot peda196 communicating with
computer 18. The call assistant 40, hearing the voice signal 16, revoices it
by
speaking the same words into the mouthpiece of the headset 38. The call
assistant's spoken words 42 are received by a speech processor system 44, to
be
4
CA 02520594 2002-02-15
24080-748D
described, which provides an editing text signal 46 to the call assistant
display 48
indicating a transcription of the call assistant's voice as well as other
control
outputs and may receive keyboard input from call assistant keyboard 50.
[025] The voice signal 16 after passing through the automatic gain
contro136 is also received by a delay circuit 21, which delays it to produce
the
delayed, modified voice signal 24 provided to the earpiece of a handset 26
used by
the deaf or hearing impaired user 14.
[026] Referring now to Fig. 2, the relay 10 may also be used with a deaf
or hearing impaired individual 14 using a personal interpreter. In this case a
voice
signal from a source proximate to the deaf or hearing impaired user 14 is
received
by a microphone 52 and relayed to the computer 18 as the voice signal 16. That
signal 16 (as buffered by recorder 19) is again received by the earpiece of
headset
38 of the call assistant 40 who revoices it as a spoken words 42.
(0271 In both the examples of Figs. I and 2, the spoken words 42 from
the call assistant 40 are received by speech processor system 44 which
produces
an editing text signal 46 separately and prior to text stream signal 20. The
editing
text signa146 causes text to appear on call assistant display 48 that may be
reviewed by the call assistant 40 for possible correction using voicing or the
keyboard 50 prior to being converted to a text stream signal 20.
[028] Referring now to Fig. 4, the relay computer 18 may be
implemented by an electronic processor 56 possibly including one or more
conventional microprocessors and a digital signal processor joined on a bus 58
with a memoryt60. The bus 58 may also communicate with various analog to
digital converters 62 providing for inputs for signals 16, 30 and 42, various
digital
to analog converters 64 providing outputs for signals 30, 24 and 17 as well as
digital I/O circuits 66 providing inputs for keyboard signal 51 and foot pedal
96
and outputs for text stream signal 20 and pre-edited editing text signal 46.
(029J Referring now to Fig. 8, the memory 60 includes a speech
recognition program 70, such as the Via Voice program manufactured by the IBM
Corporation, of a type well known in the art. The speech recognition program
70
operates under an operating system 72, such as the Windows operating system
manufactured by the Microsoft Corporation, also known in the art. The speech
recognition program 70 creates files 74 and 76 as part of its training to a
particular
5
CA 02520594 2002-02-15
24080-748D
speaker and to the text it is likely to receive. File 74 is a call assistant
specific file
relating generally to the pronunciation of the particular call assistant. File
76 is
call assistant independent and relates to the vocabulary or statistical
frequency of
word use that will be transcribed text--dependant on the pool of callers not
the call
assistant 40. File 76 will be shared among multiple call assistants in
contrast to
conventions for typical training of a speech recognition program 70, however,
file
74 will be unique to and used by only one call assistant 40 and thus is
duplicated
(not shown) for a relay having multiple call assistants 40.
(030] The memory 60 also includes program 78 of the present invention
providing for the editing features and other aspects of the invention as will
be
described below and various drivers 80 providing communication of text and
sound and keystrokes with the various peripherals described under the
operating
system 72. Memory 60 also provides a circular buffer 82 implementing recorder
19, circular buffer 84 implementing delay 21 (both shown in Fig. 1) and
circular
buffer 85 providing a queue for transcribed text prior to transmission.
Operation
of these buffers is under control of the program 78 as will be described
below.
10311 Referring now to Figs. 1 and 5, the voice signal 16 as received by
the recorder, as circular buffer 82 then. passes through a silence suppression
block
86 implemented by program 78. Generally, as voice signal 16 is received, it is
output to circular buffer 82 at a record point detennined by a record pointer
81 to
be recorded in the circular buffer 82 as a series of digital words 90. As
determined
by a playback pointer 92, these digital words 90, somewhat later in the
circular
buffer 82, are read and converted by means of digital to analog converter 64
into
headphone signal 17 communicated to headset 38. Thus, the call assistant 40
may
occasionally pause the playback of the headphone signal 17 without loss of
voice
signal 16 which is recorded by the circular buffer 82. The difference between
the
record pointer 81 and the playback pointer 92 defines the buffer fill length
94
which is relayed to the silence suppression block 86.
10321 The buffer fill length 94 may be displayed on the call assistant
display 48 shown in Fig. 3 by means of a bar graph 95 having a total width
corresponding to total size of the circular buffer 82 and a colored portion
concerning the buffer fill length 94. Altematively, a simple numerical
percentage
6
CA 02520594 2002-02-15
24080-748D
display may be provided. In this way the call assistant may keep tabs of how
far
behind she or he is in revoicing text.
[0331 The foot peda196 may be used to control movement of the
playback pointer 92 in much the same way as a conventional office dictation
unit.
While the foot pedal 96 is released, the playback pointer 92 moves through the
circular buffer 82 at normal playback speeds. When the pedal is depressed,
playback pointer 92 stops and when it is released, playback pointer 92 backs
up in
the buffer 82 by a predetennined amount and then proceeds forward at normal
playing speeds. Depression of the foot pedal 96 may thus be used to pause or
replay difficult words.
[0341 As the buffer fill length 94 increases beyond a predetermined
amount, the silence suppression block 86 may be activated to read the digital
words 90 between the record pointer 81 and playback pointer 92 to detect
silences
and to remove those silences, thus shortening the amount of buffered data and
allowing the call assistant to catch up to the conversation. In this regard,
the
silence suppression block 86 reviews the digital words 90 between the playback
pointer 92 and the record pointer 81 for those indicating an amplitude of
signal
less than a predetermined squelch value. If a duration of consecutive digital
words
90 having less than the squelch value, is found exceeding a predetermined time
limit, this silence portion is removed from the circular buffer 82 and
replaced with
a shorter silence period being the minimum necessary for clear distinction
between
words. The silence suppression block 86 then adjusts the playback pointer 92
to
reflect the shortening of the buffer fill length 94.
10351 As described above, in a preferred embodiment, the silence
suppression block 86 is activated only after the buffer fill length 94 exceeds
a
predetennined volume. However, it may alternatively be activated on a semi-
continuous basis using increasingly aggressive silence removing parameters as
the
buffer fill length 94 increases. A squelch level 98, a minimum silence period
100,
and a silence replacement value 102 may be adjusted as inputs to this silence
suppression block 86 as implemented by program 78.
[0361 Referring now to Fig. 6, after the program 78 receives the voice
signal 16 onto circular buffer 82 as indicated by process block 104, provided
the
call assistant has not depressed the peda196, the headphone signal 17 is
played
7
CA 02520594 2002-02-15
24080-748D
back as indicated by process block 106 to be received by the call assistant 40
and
revoiced as indicated by process block 108, a process outside the program as
indicated by the dotted line 109. The program 78 then connects the speech
signal
42'from the call assistant 40 to the speech recognition program 70 as
indicated by
process block 110 where it is converted to text and displayed on the call
assistant
display 48.
10371 Referring now to Fig. 3, the text is displayed within a window 112
on the call assistant display 48 and arranged into lines 114. The lines 114
organize individual text words 116 into a left to right order as in a book and
preserves a horizontal dimension of placement as the lines 114 move upward
ultimately off of the window 112 in a scrolling fashion as text is received
and
transmitted. Preserving the integrity of the lines allows the call assistant
40 to
more easily track the location of an individual word 116 during the scrolling
action.
[0381 The most recently generated text, per process block 110, is
displayed on the lowermost line 114 which forms on a word by word basis.
(039] At process block 118, the words 121 of the lowermost line are
given a first color (indicated in Fig. 3 by a lack of shading) which conveys
that
they have not yet been transmitted to the deaf or hearing impaired individual
14.
[040] At process block 120 the words are assigned an aging value
indicating how long they will be retained in a circular buffer 85 prior to
being
transmitted and hence how long they will remain the first color. The
assignment
of the aging values can be dynamic or static according to values input by the
call
assistant 40 as will be described below.
10411 As indicated by process block 122, the circular buffer 85 forms a
queue holding the words prior to transmission.
[0421 At process block 124, the words are transmitted after their
aging and this transmission is indicated changing their representation on the
display 48 to a second color 126, indicated by crosshatching in Fig. 3. Note
that
even after transmission, the words are still displayed so as to provide
continuity to
the call assistant 40 in tracking the conversation in text form.
8
CA 02520594 2002-02-15
24080-748D
10431 Prior to the words being colored the second color 126 and
transmitted (thus while the words are still in the queue 122), a cotrection of
transcription errors may occur. For example, as indicated by process block
130,
the call assistant 40 may invoke an editing routine by selecting one of the
words in
the window 112, typically by touching the word as it is displayed and
detecting
that touch using a touch screen. Altematively, the touch screen may be
replaced
with more conventional cursor control devices. The particular touched word 132
is flagged in the queue and the activation of the editing process by the touch
causes a stopping of the playback pointer 92 automatically until the editing
process is complete.
[044] Once a word is selected, the call assistant 40 may voice a
new word to replace the flagged word or type in a new word or use another
conventional text entry technique to replace the word in the queue indicated
by
process block 122. The mapping of words to spatial locations by the window 112
allows the word to be quickly identified and replaced while it is being
dynamically
moved through the queue according to its assigned aging. When the replacement
word is entered, the recorder 19 resumes playing.
[045] As an altemative to the playback and editing processes indicated by
process block 106 and 130, the call assistant 40 may enter text through a
macro
key 135 as indicated by process block 134. These macro keys 135 place
predetermined words or phrases into the queue with the touch of the macro key
135. The words or phrases may include conversational macros, such as words
placed in parentheses to indicate nonliteral context, such as (holding),
indicating
that the user is waiting for someone to come online, (sounds) indicating
nonspoken sounds necessary to understand a context, and the (unclear)
indicating
a word is not easily understood by the call assistant. Similarly, the macros
may
include call progress macros such as those indicating that an answering
machine
has been reached or that the phone is ringing. Importantly, the macros may
include conunon initial words of a sentence or phrase, such as "okay", "but",
"hello", "oh", "yes", "um", "so", "well", "no", and "bye" both to allow these
words
to be efficiently entered by the call assistant 40 without revoicing.
[046] The macro keys 135 for common initial words allow these words to
.be processed with reduced delay of the speech to text step 110 and error
cotrection
9 1
CA 02520594 2002-02-15
24080-748D
of editing process block 130. It has been found that users are most sensitive
to
delay in the appearance of these initial words and thus that reducing them
much
in~proves the comprehensibility and reduces frustration in the use of the
system.
[0471 The voice signaI received by the buffer as indicated by process
block 104 is also received by a delay line 136 implemented by circular buffer
84
and adjusted to provide delay in the voice so that the voice signal arrives at
the
caption telephone or personal interpreter at approximately the same time as
the
text. This synchronizing reduces confusion by the user.
[048] Referring now to Fig. 3, the call assistant display 48 operating
under the control of the program 78 may provide for a status indicator 138
indicating the status of the hardware in making connections to the various
users
and may include the volume control buttons 140 allowing the call assistant 40
to
independently adjust the volume of the spoken words up or down for his or her
preference. An option button 142 allows the call assistant to control the
various
parameters of the editing and speech recognition process.
[049) A DTMF button 144 allows the call assistant to directly enter
DTMF tones, for example, as may be needed for a navigation through a menu
system. Pressing of the button 144 converts the macro key 135 to a keypad on a
temporary basis.
10501 Referring now to Fig. 7, the assignment of aging of text per.process
block 120 may be functionally dependant on several.parameters. The first
parameter 146 is the location of the particular word within a block of the
conversation or sentence. It has been found that reduced delay (aging) in the
transmission of these words whether or not they are entered through the macro
process 134 or the revoicing of process block 108, decreases consumer
confusion
and frustration by reducing the apparent delay in the processing.
[0511 Error rates, as determined from the invocation of the editing
process of process block 130 may be used to also increase the aging per input
148.
As mentioned, the call assistant may control the aging through the option
button
142 shown in Fig. 3 (indicated by input 150) with inexperienced call
assistants 40
selecting for increased aging time.
10521 ' Importantly, the deaf or hearing impaired user 14 may also control
this aging time. Referring to Fig. 9, the user's terminal 22 may include, for
CA 02520594 2002-02-15
24080-748D
example, a slider control 152 providing for a range of locations between a
"faster
transcription" setting at one end and "fewer errors" setting at the other end.
Thus
the user may control the aging time to mark a preference between a few errors
but
faster transcription or much more precise transcription at the expense of some
delay.
10531 It will be understood that the mechanisms described above may
also be realized in collections of discrete hardware rather than in an
integrated
electronic computer according to methods well known in the art.
[054] It should be noted that the present invention provides utility even
against the expectation of increased accuracy in computer speech recognition
and
it is therefore considered to cover applications where the call assistant may
perform no or little revoicing while using the editing mechanisms described
above
to correct for machine transcription errors.
(055] It will be understood that the digital tape recorder 19, including the
foot pedal 96 and the silence suppression block 86 can be equally used with a
conventional relay in which the call assistant 40 receiving a voice signal
through
the headset 38 types, rather than revoices, the signal into a conventional
keyboard
50. In this case the interaction of the digital tape recorder 19 and the
editing
process may be response to keyboard editing commands (backspace etc) rather
than the touch screen system described above. A display may be used to provide
the bar graph 95 to the same purposes as that described above...
10561 It is specifically intended that the present invention not be limited
to the embodiments and illustrations contained herein, but that modified forms
of
those embodiments including portions of the embodiments and combinations of
elements of different embodiments also be included as come within the scope of
the following claims.
11
