Conference system suitable for high-noise environment

Abstract

The invention provides a conference system suitable for a high-noise environment. The conference system comprises a processing end, a first conference terminal and a second conference terminal, the processing end is used for acquiring voice signals continuously sent by the first conference terminal; intercepting a sample voice signal of a preset duration from the voice signal every preset time interval, and calculating a signal-to-noise ratio of a useful signal to background noise in the sample voice signal; if the signal-to-noise ratio of the sample voice signal is smaller than a preset threshold value, performing noise reduction processing on the voice signal to obtain a pure voice signal; inputting the voice features of the pure voice signal into a decoder comprising an acoustic model and a language model to obtain a pure voice text; and sending the pure voice signal and the pure voice text to the second conference terminal, so that the second conference terminal plays the pure voice signal and / or displays the pure voice text. According to the invention, the voice signal collected by the terminal can be subjected to noise reduction processing, and interference of noise to a conference is reduced.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a conference system suitable for high-noise environments. Background Art

[0002] With the rapid development of Internet technology, Internet concepts are applied to all walks of life and are gradually changing the way we live and work.

[0003] At present, many online conference system software have been developed. You can log in to the online conference system using a computer or a smart phone. Smart phones have strong flexibility and can start meetings anytime and anywhere through the Internet. This brings convenience to work, but also creates some problems. The outdoor environment is affected by weather factors, traffic factors, etc., and may generate certain noise during the meeting, interfering with the meeting. Therefore, there is an urgent need to develop an online conference system suitable for noisy environments. Summary of the invention

[0004] In view of this, the purpose of the embodiments of the present application is to provide a conference system suitable for high-noise environments, which can perform noise reduction processing on voice signals collected by terminals and reduce the interference of noise on conferences.

[0005] In a first aspect, an embodiment of the present application provides a conference system suitable for a high noise environment, the conference system comprising a processing end, a first conference terminal, and a second conference terminal; the processing end is used to:

[0006] Acquire the voice signal continuously sent by the first conference terminal;

[0007] At every preset time interval, a sample voice signal of a preset duration is intercepted from the voice signal, and a signal-to-noise ratio of a useful signal to background noise in the sample voice signal is calculated;

[0008] If the signal-to-noise ratio of the sample speech signal is less than a preset threshold, the speech signal is subjected to noise reduction processing to obtain a pure speech signal;

[0009] Inputting the speech features of the clean speech signal into a decoder including an acoustic model and a language model to obtain a clean speech text;

[0010] The clean voice signal and the clean voice text are sent to the second conference terminal, so that the second conference terminal plays the clean voice signal and / or displays the clean voice text.

[0011] In a possible implementation, calculating the signal-to-noise ratio of the useful signal to the background noise in the sample speech signal includes: calculating the signal-to-noise ratio SNR according to the formula SNR=Ps / Pt, wherein Ps represents the average power of the useful signal and Pt represents the average power of the background noise.

[0012] In a possible implementation manner, if the first conference terminal is located in an environment where noise and human voice appear at intervals; the performing noise reduction processing on the voice signal to obtain a pure voice signal includes:

[0013] Extracting a first speech feature of the speech signal;

[0014] The similarity between the first speech feature and a preset human voice feature is calculated, and the clean speech signal is generated based on the speech features whose similarity is greater than a preset similarity threshold.

[0015] In a possible implementation manner, if noise and human voice are mixed in the environment where the first conference terminal is located; the performing noise reduction processing on the voice signal to obtain a pure voice signal further includes:

[0016] Segmenting the speech signal and performing a fast Fourier transform, so that the speech signal is converted from a time domain signal to a frequency domain signal;

[0017] The transformed frequency domain signal is input into a low-pass filter to filter out the noise signal in the high-frequency bandwidth and obtain a pure speech signal.

[0018] In a possible implementation, inputting the speech features of the clean speech signal into a decoder including an acoustic model and a language model to obtain a clean speech text includes:

[0019] Extracting a second speech feature of the clean speech signal;

[0020] Inputting the second speech feature into the acoustic model to obtain a speech word;

[0021] The phonetic words are input into the language model to obtain the clean phonetic text.

[0022] In a possible implementation, inputting the second speech feature into the acoustic model to obtain a speech word includes:

[0023] Using a hidden Markov model or a mixed Gaussian model, converting the input second speech feature into a phoneme feature, and obtaining a phoneme corresponding to the phoneme feature;

[0024] A pronunciation dictionary is used to map interrelated phonemes to corresponding phonetic words.

[0025] In a possible implementation, the step of inputting the speech word into the language model to obtain the clean speech text includes:

[0026] According to preset grammatical rules, the probability of the phonetic words appearing in a complete sentence is calculated to obtain a complete sentence consisting of the phonetic words with a maximized probability, and the pure phonetic text is composed of multiple consecutive complete sentences.

[0027] In a possible implementation, the processing end is further used to: input the speech signal into a deep residual shrinkage network with speech recognition function to obtain a pure speech text.

[0028] In a possible implementation manner, the first conference terminal is further configured to send the noisy voice signal to the second conference terminal;

[0029] If the conference mode is a scene display mode, the second conference terminal is used to play the voice signal containing noise and display the pure voice text;

[0030] If the conference mode is a voice interaction mode, the second conference terminal is used to play the clean voice signal and / or display the clean voice text.

[0031] The embodiment of the present application provides a conference system suitable for a high-noise environment, which can perform noise reduction processing on voice signals collected by a terminal to reduce the interference of noise on the conference.

[0032] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A schematic diagram of the structure of a conference system suitable for a high noise environment provided in an embodiment of the present application is shown;

[0035] Figure 2 A flow chart for processing a speech signal provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0036] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0037] With the rapid development of Internet technology, Internet concepts are applied to all walks of life and are gradually changing the way we live and work.

[0038] At present, many online conference system software have been developed. You can log in to the online conference system using a computer or a smart phone. Smart phones have strong flexibility and can start meetings anytime and anywhere through the Internet. This brings convenience to work, but also creates some problems. The outdoor environment is affected by weather factors, traffic factors, etc., and may generate certain noise during the meeting, interfering with the meeting. Therefore, there is an urgent need to develop an online conference system suitable for noisy environments.

[0039] Based on the above problems, an embodiment of the present application provides a conference system suitable for a high-noise environment, which can perform noise reduction processing on the voice signal collected by the terminal to reduce the interference of noise on the conference.

[0040] The technical solutions in the present application will be clearly and completely described below in conjunction with the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0041] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0042] See also Figure 1 As shown, Figure 1A structural diagram of a conference system suitable for a high-noise environment provided in an embodiment of the present application, the conference system 100 includes a processing terminal 101, a first conference terminal 102, and a second conference terminal 103. It should be noted that the first conference terminal (the second conference terminal) can be one or more. The first conference terminal and the second conference terminal can be any electronic device with the functions of recording sound, playing sound, and displaying text, such as a smart phone, a tablet computer, and a laptop computer. "First" and "second" are used to distinguish between the terminal that is recording the voice signal and the terminal that is playing the voice signal, and the first conference terminal and the second conference terminal can be converted to each other.

[0043] See also Figure 2 As shown, Figure 2 A flowchart of processing a voice signal is provided in an embodiment of the present application, and the processing flow of the voice signal by the processing end includes:

[0044] S201: Acquire a voice signal continuously sent by the first conference terminal.

[0045] Users can use the first conference terminal to transmit voice signals to the second conference terminal at any location, such as indoors or outdoors. Since the environment of the first conference terminal may be disturbed by external noise, the processing end needs to first perform noise reduction and voice-to-text processing on the voice signal collected by the first conference terminal, and then the processing end sends the processed voice signal and voice text to the second conference terminal. Therefore, the first conference terminal needs to continuously send the collected voice signal to the processing end.

[0046] S202: every preset time interval, extract a sample speech signal of a preset duration from the speech signal, and calculate a signal-to-noise ratio of a useful signal to background noise in the sample speech signal.

[0047] In the embodiment of the present application, the processing end can directly perform noise reduction processing on the received voice signal, but some voice signals do not contain noise, or contain less noise. Even if such voice signals are not subjected to noise reduction processing, they will not affect semantic understanding and will not interfere with the progress of the meeting. If all voice signals are subjected to noise reduction processing, the workload of the processing end will inevitably increase, resulting in a waste of resources on the processing end. Therefore, step S202 is used to pre-judge the voice signal. At regular intervals, a segment of voice signal is intercepted as a sample voice signal, and the sample voice signal is used to represent the signal-to-noise situation of the voice signal. The signal-to-noise ratio of this segment of the sample voice signal is calculated, and based on the size of the calculated signal-to-noise ratio, it is determined whether to perform noise reduction processing on the voice signal. The two parameters of the time interval and the preset duration are set according to the actual situation. If the accuracy of the conference call is required to be high, the time interval can be set to a smaller value and the preset duration can be set to a larger value.

[0048] For example, every 5 minutes, a 10-second sample speech signal is intercepted from the speech signal, and the signal-to-noise ratio of the 10-second sample speech signal is calculated. If the signal-to-noise ratio is large, it means that the sample speech signal has less noise, which means that the speech signal has less noise and there is no need to perform noise reduction processing on the speech signal.

[0049] S203: If the signal-to-noise ratio of the sample speech signal is less than a preset threshold, perform noise reduction processing on the speech signal to obtain a pure speech signal.

[0050] A threshold is preset according to the actual situation. If the signal-to-noise ratio is greater than or equal to the preset threshold, it means that the sample speech signal has less noise, that is, the speech signal has less noise, and there is no need to perform noise reduction on the speech signal. Otherwise, the speech signal needs to be processed for noise reduction. Noise includes all non-human sounds such as wind, rain, whistle, and mechanical operation. Alternatively, the type of noise can be adjusted according to the content of the meeting. For example, in an on-site meeting related to the performance of automobile brakes, the sound of car driving and car braking is no longer noise, and the type of noise can be adjusted according to the actual situation. After removing the useless noise in the speech signal, a pure speech signal is obtained.

[0051] S204: Input the speech features of the clean speech signal into a decoder including an acoustic model and a language model to obtain a clean speech text.

[0052] The acoustic model (AM) takes speech features as input, calculates the probability between the audio and corresponding phonemes, and converts the sound into the corresponding phonemes through the determination of the probability threshold. Then, based on the correspondence between phonemes and words, the phonemes are converted into words, thereby establishing an acoustic model.

[0053] The language model (LM) mainly integrates grammar and word knowledge. The language model calculates the probability of a word appearing in a sentence, or the probability of several words forming a sentence, forms sentences through recognition and judgment, and finally obtains a text composed of sentences.

[0054] In this embodiment, we connect the acoustic model and the language model through a decoder, input the speech features of the clean speech signal into the decoder, and the acoustic model and the language model connected to the decoder continue to process the speech features of the clean speech signal to obtain a clean speech text.

[0055] S205: Send the clean voice signal and the clean voice text to the second conference terminal, so that the second conference terminal plays the clean voice signal and / or displays the clean voice text.

[0056] The second conference terminal that establishes a conference relationship with the first conference terminal can be one or more. The clean voice signal and the clean voice text are sent to the second conference terminal so that the second conference terminal plays the clean voice signal, or displays the clean voice text in the form of adding subtitles.

[0057] The conference system suitable for high-noise environments provided in the embodiments of the present application can perform noise reduction processing on voice signals collected by the terminal to reduce the interference of noise on the conference.

[0058] In step S202, the calculating of the signal-to-noise ratio of the useful signal to the background noise in the sample speech signal includes: calculating the signal-to-noise ratio SNR according to the formula SNR=Ps / Pt, wherein Ps represents the average power of the useful signal and Pt represents the average power of the background noise.

[0059] The signal-to-noise ratio is defined as the ratio of the power of the signal (meaningful input) to the power of the background noise (meaningless or unwanted input). Where Ps and Pt are both average powers. It is important to note that the signal and noise powers must be measured at the same or equivalent points in the system and within the same system bandwidth.

[0060] In step S203, the speech signal is subjected to noise reduction processing to obtain a pure speech signal. The present application embodiment provides two noise reduction methods, which are respectively applicable to different application scenarios.

[0061] Noise reduction method 1: If the first conference terminal is located in an environment where noise and human voice appear at intervals, extract the first voice feature of the voice signal; calculate the similarity between the first voice feature and the preset human voice feature, and generate the pure voice signal based on the voice features whose similarity is greater than the preset similarity threshold.

[0062] When noise and human voice appear in the environment where the first conference terminal is located, the voice signal is segmented according to a certain time interval. Some of the voice signals in the segments are human voices and some are noise. By classifying the voice signals, it is determined which time periods belong to human voice signals and which time periods belong to noise signals, thereby realizing the classification of human voice signals and noise signals, and further realizing noise reduction processing of the voice signals.

[0063] First, the speech signal is preprocessed, including framing, windowing and other operations, to divide the speech data into very small segments, and obtain continuous multi-frame speech segments, which can ensure easy processing while minimizing signal distortion. Then, the frequency domain feature is obtained by Fourier transform for each frame signal, and the obtained frequency domain feature is determined as the first speech feature. The frequency domain feature can be set as MFCC (Mel Frequency Cepstral Coefficient) feature, etc.

[0064] Calculate the distance between the first speech feature and the preset vocal feature, such as Euclidean distance, Manhattan distance, Chebyshev distance, average, median, maximum, minimum, etc. of multiple distances, and use the calculated distance as the similarity. If the similarity of the first speech feature is greater than the similarity threshold, classify the first speech feature into a vocal feature, and generate a pure speech signal based on the multiple vocal features obtained by classification.

[0065] Noise reduction method 2: If the first conference terminal is located in an environment where noise and human voices are mixed, the voice signal is segmented and fast Fourier transformed, and the voice signal is converted from a time domain signal to a frequency domain signal; the converted frequency domain signal is input into a low-pass filter to filter out the noise signal of the high-frequency bandwidth to obtain a pure voice signal.

[0066] Unlike the first noise reduction method, in which noise and human voice are separated from each other, the second noise reduction method is suitable for situations where human voice and noise are mixed together. In speech signal processing, the speech signal needs to be segmented and subjected to fast Fourier transform (FFT). Because there is noise in the time domain, a certain bandwidth appears in the frequency. At this time, a low-pass filter can be used to filter out the high-frequency components from the signal, thereby achieving the effect of denoising.

[0067] In step S204, the speech features of the clean speech signal are input into a decoder including an acoustic model and a language model to obtain a clean speech text, including:

[0068] S2041, extracting the second speech feature of the clean speech signal; S2042, inputting the second speech feature into the acoustic model to obtain speech words; S2043, inputting the speech words into the language model to obtain the clean speech text.

[0069] First, the clean speech signal is preprocessed, including framing, windowing and other operations, to divide the speech into very small segments, and obtain multiple continuous frames of clean speech segments, which can ensure easy processing while minimizing signal distortion. Then, the frequency domain features can be obtained by Fourier transforming each frame of the signal, and the obtained frequency domain features are determined as the second speech features.

[0070] The decoder connects the acoustic model and the language model. The acoustic model realizes the conversion from speech features (continuous numbers) to speech words (the smallest unit of text). The language model splices independent speech words into speech text that meets grammatical requirements and conforms to the language expression habits of ordinary people. The extracted second speech feature is the input of the decoder, and the pure speech text is the output of the decoder.

[0071] In step S2042, the second speech feature is input into the acoustic model to obtain a speech word, including: using a hidden Markov model or a mixed Gaussian model to convert the input second speech feature into a phoneme feature, and obtaining the phoneme corresponding to the phoneme feature; using a pronunciation dictionary to map mutually related phonemes to corresponding speech words.

[0072] Phoneme is the smallest unit of pronunciation. Phoneme is text, and phoneme feature is a vector. There is a mapping relationship between phoneme and phoneme feature. It is necessary to use hidden Markov model or mixed Gaussian model to convert the second speech feature into phoneme feature. Based on the mapping relationship between preset phoneme feature and phoneme, multiple phonemes corresponding to the second speech feature are obtained, and then the phoneme string is mapped to the corresponding speech word with the help of pronunciation dictionary. Here, "phoneme string" refers to multiple consecutive related phonemes.

[0073] In step S2043, the voice words are input into the language model to obtain the clean voice text, including: according to preset grammatical rules, the probability of the voice words appearing in a complete sentence is calculated, and a complete sentence composed of the voice words with maximized probability is obtained, and the clean voice text is composed of multiple consecutive complete sentences.

[0074] The speech words output by the acoustic model are independent of each other. The speech words need to be spliced ​​into sentences, and then the sentences are spliced ​​into text. Assuming that the speech words can be spliced ​​into N sentences, the probability of the speech word belonging to each of the N sentences is calculated based on the existing grammatical rules, common word collocation habits, and the speaking habits of ordinary people. The sentence with the largest probability value is taken as the sentence to which the speech word belongs, and multiple complete sentences are obtained. Then, multiple complete sentences are spliced ​​into a complete pure speech text.

[0075] Furthermore, the processing end is also used to: input the voice signal into a deep residual shrinkage network with a voice recognition function to obtain a pure voice text.

[0076] The deep residual shrinkage network is a method that integrates signal denoising, feature extraction and pattern recognition. The deep residual shrinkage network can automatically eliminate information irrelevant to the current task through a soft thresholding layer with an adaptive threshold, and achieve accurate pattern recognition of strong noise data. It can be applied in the field of speech recognition. Its working principle can be understood as: paying attention to unimportant features through the attention mechanism and setting them to zero through the soft threshold function; or, paying attention to important features through the attention mechanism and retaining them, thereby enhancing the ability of deep neural networks to extract useful features from noisy signals.

[0077] When classifying samples, there will inevitably be some noise in the samples, such as Gaussian noise, pink noise, Laplace noise, etc. More generally speaking, the samples are likely to contain information that is irrelevant to the current classification task, which can also be understood as noise. These noises may have an adverse effect on the classification effect.

[0078] For example, when a meeting is held on the roadside or on a train, the voices of the participants may be mixed with the sound of vehicle horns and wheels. When these sound signals are used for speech recognition, the recognition effect will inevitably be affected by the horns and wheels. From the perspective of deep learning, the features corresponding to these horns and wheels should be deleted inside the deep neural network to avoid affecting the effect of speech recognition.

[0079] The embodiment of the present application adopts a deep residual shrinkage network with speech recognition function to reduce the noise of the speech signal and convert the speech signal into pure speech text.

[0080] Furthermore, the first conference terminal is also used to send the noisy voice signal to the second conference terminal; if the conference mode is the scene display mode, the second conference terminal is used to play the noisy voice signal and display the clean voice text. If the conference mode is the voice interaction mode, the second conference terminal is used to play the clean voice signal and / or display the clean voice text.

[0081] In the embodiment of the present application, the conference mode includes a scenario display mode and a voice interaction mode. In the scenario display mode, it is necessary to display the on-site environment where the first conference terminal is located, for example, displaying the construction environment, construction progress, traffic flow, etc. At this time, the voice signal collected by the first conference terminal is subjected to noise reduction processing, and the on-site environment cannot be restored 100%. Therefore, the second conference terminal plays the noisy voice signal and displays the pure voice text in the form of subtitles to compensate for the impact of noise on the meeting. In the voice interaction mode, there is no need to display the on-site environment, only the voice content needs to be displayed. At this time, the second conference terminal plays the pure voice signal, or displays the pure voice text, or displays the pure voice text while playing the pure voice signal.

[0082] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-mentioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-mentioned embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A conference system suitable for high noise environment, characterized in that: The conference system includes a processing end, a first conference terminal, and a second conference terminal; the processing end is used for: Acquire the voice signal continuously sent by the first conference terminal; At every preset time interval, a sample voice signal of a preset duration is intercepted from the voice signal, and a signal-to-noise ratio of a useful signal to background noise in the sample voice signal is calculated; If the signal-to-noise ratio of the sample speech signal is less than a preset threshold, the speech signal is subjected to noise reduction processing to obtain a pure speech signal; Inputting the speech features of the clean speech signal into a decoder including an acoustic model and a language model to obtain a clean speech text; The clean voice signal and the clean voice text are sent to the second conference terminal, so that the second conference terminal plays the clean voice signal and / or displays the clean voice text.

2. The conference system suitable for high noise environment according to claim 1, characterized in that: The calculating of the signal-to-noise ratio of the useful signal to the background noise in the sample speech signal includes: calculating the signal-to-noise ratio SNR according to the formula SNR=Ps / Pt, wherein Ps represents the average power of the useful signal and Pt represents the average power of the background noise.

3. The conference system suitable for high noise environment according to claim 1, characterized in that: If noise and human voice appear in the environment where the first conference terminal is located, the noise reduction processing on the voice signal to obtain a pure voice signal includes: Extracting a first speech feature of the speech signal; The similarity between the first speech feature and a preset human voice feature is calculated, and the clean speech signal is generated based on the speech features whose similarity is greater than a preset similarity threshold.

4. The conference system suitable for high noise environment according to claim 1, characterized in that: If noise and human voice are mixed in the environment where the first conference terminal is located; the noise reduction processing of the voice signal to obtain a pure voice signal also includes: Segmenting the speech signal and performing a fast Fourier transform, so that the speech signal is converted from a time domain signal to a frequency domain signal; The transformed frequency domain signal is input into a low-pass filter to filter out the noise signal in the high-frequency bandwidth and obtain a pure speech signal.

5. The conference system suitable for high noise environment according to claim 1, characterized in that: The step of inputting the speech features of the clean speech signal into a decoder including an acoustic model and a language model to obtain a clean speech text comprises: Extracting a second speech feature of the clean speech signal; Inputting the second speech feature into the acoustic model to obtain a speech word; The phonetic words are input into the language model to obtain the clean phonetic text.

6. The conference system suitable for high noise environment according to claim 5, characterized in that: The step of inputting the second speech feature into the acoustic model to obtain a speech word comprises: Using a hidden Markov model or a mixed Gaussian model, converting the input second speech feature into a phoneme feature, and obtaining a phoneme corresponding to the phoneme feature; A pronunciation dictionary is used to map interrelated phonemes to corresponding phonetic words.

7. The conference system suitable for high noise environment according to claim 5, characterized in that: The step of inputting the speech word into the language model to obtain the clean speech text includes: According to preset grammatical rules, the probability of the phonetic words appearing in a complete sentence is calculated to obtain a complete sentence consisting of the phonetic words with a maximized probability, and the pure phonetic text is composed of multiple consecutive complete sentences.

8. The conference system suitable for high noise environment according to claim 1, characterized in that: The processing end is also used to: input the voice signal into a deep residual shrinkage network with a voice recognition function to obtain a pure voice text.

9. The conference system suitable for high noise environment according to claim 1, characterized in that: The first conference terminal is further used to send the noisy voice signal to the second conference terminal; If the conference mode is a scene display mode, the second conference terminal is used to play the voice signal containing noise and display the pure voice text; If the conference mode is a voice interaction mode, the second conference terminal is used to play the clean voice signal and / or display the clean voice text.

Images