GB/T 36464.1-2020Information technology - Intelligent speech interaction system - Part 1: General specifications (English PDF)

3.1 Voice interaction

Information transmission and communication activities between humans and functional units through voice.

[GB/T 36464.2-2018, definition 3.1]

3.2 Voice interaction system

A system that is composed of functional units (or combinations thereof), data resources, etc. that can realize voice interaction with humans.

[GB/T 36464.2-2018, definition 3.2]

3.3 Intelligent voice interaction system

Based on all or part of artificial intelligence technologies such as speech recognition, semantic understanding, and speech synthesis, it is composed of intelligent software and hardware, and has intelligent

A voice interaction system with human-computer interaction capabilities.

3.4 Human-computer interaction

Between humans and functional units, in order to complete certain tasks, information transmission and communication activities are carried out in a certain interactive manner.

3.5 Functional unit

A hardware entity, or a software entity, or a hardware entity and a software entity that can complete a specific task.

[GB/T 5271.1-2000, definition 01.01.40]

3.6 Speech synthesis

The process of synthesizing human language through mechanical and electronic methods.

[GB/T 21024-2007, definition 3.1]

3.30 Affective computing

The collection, recognition, decision-making and expression of specific emotions in the process of human-computer interaction.

7.1 Voice wake up

7.1.1 Command word wake up

The system should support the use of predefined command words to wake up the system by voice.

7.1.2 Command word voiceprint wake up

In the process of voice wake-up, the system should support the use of text-related voiceprint recognition and command word matching, and it can wake up after the voiceprint is confirmed successfully system.

7.1.3 Custom wake-up command word

The system should support customization of the command word used for voice wake-up.

7.1.4 Multiple wake-up command words

The system should support the use of different command words for voice wake-up; it can enter the corresponding state or mode according to the specified wake-up command word.

7.1.5 Multi-audio stream monitoring

When the system is waking up by voice, it should support simultaneous monitoring of multiple audio streams.

7.2 Sound source localization

The system should support the location of the sound source by calculating the plane angle, azimuth angle and distance of the sound source.

7.3 Voiceprint recognition

7.3.1 General requirements

The system should support the following voiceprint recognition functions.

a) Text-related voiceprint recognition;

b) Text-independent voiceprint recognition;

c) Voiceprint recognition of specified text;

d) Voiceprint model training;

e) Voiceprint model adaptation;

f) Voiceprint confirmation;

g) Voiceprint recognition;

h) Voiceprint detection;

i) Voiceprint tracking;

j) Language-related voiceprint recognition;

k) Language-independent voiceprint recognition.

The above functional descriptions and requirements shall meet the requirements of Chapter 3 of SJ/T 11380-2008.

7.3.2 Voiceprint text acquisition

The system should support the acquisition of specified text or custom text for voiceprint model training, voiceprint model adaptation, voiceprint confirmation and voiceprint identify.

7.4 Speech enhancement

7.4.1 Noise suppression

The system should support the suppression of background noise in the input speech and improve the signal-to-noise ratio of the speech.

7.4.2 Reverberation cancellation

The system should support the suppression of late reverberation in the input speech to improve the clarity and intelligibility of the speech signal.

7.5 Format conversion

The system should support the conversion of audio format to another format to meet the requirements of voice processing.

7.6 Resampling

The system should support changing the sampling rate of digital voice signals to meet the requirements of voice processing.

8.1 Speech recognition

8.1.1 General requirements

The system should support all or most of the following speech recognition functions.

a) Chinese speech recognition service;

b) Multilingual recognition;

c) Multi-dialect recognition;

d) Multilingual mixed reading recognition;

e) Custom syntax;

f) Personalized identification;

g) Multiple candidates for recognition results;

h) Custom hot words;

i) Advanced recognition results;

j) Language information recognition;

k) Speaker information recognition.

The above functional descriptions and requirements should meet the requirements of 4.2 and 4.3 in GB/T 34083-2017.

8.1.2 Voice recognition method

The system should support near-field audio processing and/or far-field audio processing; it should at least support keyword recognition, command word recognition, and continuous

A type of speech recognition.

8.2 Semantic understanding

8.2.1 Custom Semantic Dictionary

The system can support application-defined semantic dictionaries and user-defined semantic dictionaries.

8.2.2 Custom semantic library

The system can support application-defined semantic libraries and user-defined semantic libraries.

8.2.3 Fuzzy recognition

The system should correctly deal with typos, synonyms, and more and fewer words.

8.2.4 Semantic extraction

In the interaction process of the system, the semantic elements and the key intent of the user should be extracted.

8.2.5 Semantic Sorting

The system can provide multiple sorted comprehension results in the semantic comprehension results for users to choose or confirm again.

8.3 Speech synthesis

The system should support all or most of the following speech synthesis functions.

a) Chinese speech synthesis;

b) Streaming speech synthesis;

c) Multiple synthetic text encodings;

d) Personalized synthesis;

e) Multilingual synthesis;

f) Multi-dialect synthesis;

g) Multilingual mixed reading synthesis;

h) Synthetic audio multi-timbre;

i) User-defined word segmentation;

j) User-defined pronunciation;

k) Synthetic text location information;

l) Text segmentation and pinyin information;

m) Audio time information.

The above functional descriptions and requirements should meet the requirements of 4.2 and 4.3 in GB/T 34145-2017.

8.4 Endpoint detection

8.4.1 Single Endpoint Detection

The system should support detecting the start point and end point of the first speech segment from a continuous audio stream.

8.4.2 Multi-endpoint detection

The system should support detecting the start and end points of multiple speech segments from a continuous audio stream.

8.4.3 Endpoint detection sensitivity setting

The system should support setting the voice waiting timeout period and tail mute length, and adjust the sensitivity of voice endpoint detection.

8.5 Voice codec

8.5.1 Variable rate encoding

The system should support changing the code stream rate of the coded speech output by the speech coding algorithm by setting the coding level or by other means.

8.5.2 Compression level setting

The system should support setting the compression level of the speech coding algorithm according to the current network conditions, system performance and other requirements.

8.6 Full duplex interaction

The system should support full-duplex voice interaction; in this state, it should support one-time voice wake-up and voice interruption at any time, enabling contextual language

Context and open dialogue management can control the rhythm of dialogue and predict user intentions.

8.7 Emotional Computing

The system should support emotional calculation using voice signals as the carrier.

A.1 Overview

This appendix gives some parameter definitions and calculation methods used to describe the intelligent voice interaction system.

A.2 Pickup distance

When the distance between the sound source and the pickup device is less than or equal to 1m, it is the near field; when the distance between the sound source and the pickup device is > 1m, it is the far field.

A.3 Voice interaction success rate

In a certain period of time, the total number of successful voice interactive sessions accounted for the percentage of the total number of effective voice interactive sessions. "Successful Voice

"Interactive session" refers to a voice interactive session in which a complete voice service result is obtained without errors during the period; "effective voice interactive session" refers to

All voice interactive sessions remove failed sessions due to user terminal failure or user behavior or parameter errors.

Refer to formula (A.1) for the calculation method of interaction success rate.

PS=

SF×

100% (A.1)

Where.

PS---Interaction success rate, %;

S ---the number of successful interactions;

F ---The number of failed interactions.

A.4 Voice wakeup

A.4.1 Wake-up rate

The ratio of the number of successful wake-ups to the total number of voice wake-ups in a voice wake-up operation within a certain period of time. Used to describe the voice wake-up operation

For the correct response, see formula (A.2) for the calculation method.

rhosw=

Nsw

Nw ×

100% (A.2)

Where.

rhosw --- arousal rate, %;

Nsw---successful wake-up times;

Nw ---The number of voice wake-up operations.

A.4.2 Frequency of false wakeups

The frequency of false wakeup describes the frequency of false wakeup operations in unit time, and the calculation method is shown in formula (A.3).

fFW=

NFW (A.3)

Where.

fFW --- False wake-up frequency, the unit is times per hour (times/h);

NFW---the number of false awakenings during the investigation period;

T --- the duration of the evaluation, in hours (h).

A.5 Speech recognition

A.5.1 Word accuracy

For the calculation method of word accuracy, please refer to 5.2.1 of GB/T 21023-2007.

A.5.2 Sentence recognition rate

The calculation method of sentence recognition rate is shown in formula (A.4).

Psr=

Nsr

Nsi×

100% (A.4)

Where.

Psr---sentence recognition rate, %;

Nsr---the number of sentences correctly recognized by the intelligent voice interaction system;

Nsi---mark the total number of sentences.

A.5.3 Initial response time

The time elapsed from when the user's valid voice input is detected to when the first part of the recognition result is obtained, in milliseconds (ms), used to describe the real-time response of speech recognition.

A.5.4 End response time

The time elapsed from when the user's valid voice input is detected to when the last part of the recognition result is obtained, in milliseconds

Seconds (ms), used to describe the real-time response of speech recognition.

A.6 Semantic understanding

A.6.1 Correct rate of semantic understanding

See formula (A.5) for the calculation method of semantic understanding correct rate.

RSS=

NSS N ×100% (A.5)

Where.

RSS---The correct rate of semantic understanding, %;

NSS---the number of times the operation intention and semantic elements are correctly judged;

N ---The total number of times the user input is correctly recognized text information.

A.6.2 Correct response rate

Refer to formula (A.6) for the calculation method of the correct rate of Chinese comprehension response.

A.9 Voiceprint recognition rate

The voiceprint recognition rate includes parameters such as false rejection rate, false acceptance rate, missed recognition rate, and false alarm rate. Refer to SJ/T 11380- for the calculation method.

The provisions of 3.3.2 in.2008.

A.10 Voice coding and decompression rate

The speech coding/decompression ratio is the ratio of the code stream rate of the compressed audio output by the speech compression algorithm to the input audio to be compressed.

A.11 Speech enhancement

A.11.1 Signal-to-noise ratio improvement

The signal-to-noise ratio improvement is the ratio of the output speech signal-to-noise ratio of the speech enhancement function unit to the input speech signal-to-noise ratio.

A.11.2 Noise suppression

Refer to formula (A.7) for the calculation method of noise suppression.

DNR=10log

N-1 n=0

|nuin(n)|2

N-1 n=0

|nuout(n)|2 (A.7)

Where.

DNR --- the amount of noise suppression, in decibels (dB);

nuin(n)---the amplitude of the nth noise signal in the input signal;

nuout(n)---the amplitude of the nth noise signal in the output signal;

N --- the total number of frequency components of the input signal spectrum.

A.12 Sound source location

A.12.1 Plane angle positioning error

The plane angle positioning error is the difference between the plane angle of the sound source position calculated by the sound source positioning function unit and the true value.

A.12.2 Pitch angle positioning error

The pitch angle positioning error is the difference between the pitch angle of the sound source position calculated by the sound source positioning function unit and the true value.

A.12.3 Distance positioning error

The distance positioning error is the difference between the sound source position distance calculated by the sound source positioning function unit and the true value.

A.13 Voice interruption success rate

In dialogue management, the success rate of voice interruption refers to the ratio of the number of correct responses to the voice interruption operation in a certain period of time. its […]