Speech recognition method, speech recognition model, speech recognition device and storage medium

Abstract

The invention provides a speech recognition method, a speech recognition model, a speech recognition device and a storage medium. The speech recognition method comprises the following steps: acquiring a speech waveform to be recognized; inputting the feature vector of the voice waveform into a routing network, so that the routing network screens out at least one target voice recognition network corresponding to the feature vector from a plurality of voice recognition networks; inputting the feature vector of the voice waveform into at least one target voice recognition network to obtain at least one voice recognition result; and determining a final recognition result of the voice waveform according to the at least one voice recognition result. According to the embodiment of the invention, the voice recognition effect can be improved under the condition of considering the reasoning speed.

Description

Technical Field

[0001] This application relates to the field of speech recognition technology, specifically to a speech recognition method, speech recognition model, device, and storage medium. Background Technology

[0002] In recent years, Automatic Speech Recognition (ASR) technology has made significant progress thanks to its end-to-end neural network architecture. Typical models include Connectionist Temporal Classification (CTC), attention mechanisms, and Transformer structures. However, in multilingual or low-resource scenarios, these methods still face the following bottlenecks: Strong data dependency: End-to-end models typically require large-scale, high-quality labeled data, which is difficult to meet in low-resource languages. Limited scalability: Adding a new language or task requires overall retraining or significant fine-tuning, resulting in high computational and storage overhead. Insufficient utilization of expert knowledge: Traditional models treat all inputs equally, making it difficult to activate specific language or acoustic experts as needed, leading to parameter redundancy and low inference efficiency. Summary of the Invention

[0003] In view of this, this application proposes a speech recognition method, a speech recognition model, a device, and a storage medium.

[0004] The first aspect of this application proposes a speech recognition method, including:

[0005] Obtain the waveform of the speech to be recognized;

[0006] The feature vector of the speech waveform is input into the routing network so that the routing network selects at least one target speech recognition network corresponding to the feature vector from multiple speech recognition networks;

[0007] The feature vectors of the speech waveform are respectively input into the at least one target speech recognition network to obtain at least one speech recognition result;

[0008] The final recognition result of the speech waveform is determined based on the at least one speech recognition result.

[0009] This application embodiment selects at least one target speech recognition network corresponding to the feature vector from multiple speech recognition networks, and uses at least one target speech recognition network to determine the final recognition result of the speech waveform, which can improve the overall speech recognition performance while taking speed into account.

[0010] In this embodiment of the application, after acquiring the speech waveform to be recognized, the method further includes:

[0011] The original acoustic features are obtained by extracting features from the speech waveform;

[0012] The original acoustic features are encoded into high-dimensional features with contextual semantic information, and the high-dimensional features are used as the feature vector of the speech waveform.

[0013] In this embodiment, each speech recognition network includes an encoder and a decoder; the feature vectors of the speech waveform are respectively input into the at least one target speech recognition network to obtain at least one speech recognition result, including:

[0014] For any target speech recognition network, a second sequence corresponding to the target encoder of the target speech recognition network is determined from a first sequence; the first sequence includes multiple encoders stacked in a preset order, and the multiple encoders correspond one-to-one with the multiple speech recognition networks; the second sequence refers to the serial sequence in the first sequence with the target encoder as the tail encoder.

[0015] The feature vector of the speech waveform is input into the second sequence to obtain the target encoding result of the target encoder; the encoding result output by the previous encoder in the second sequence is used as the input of the next encoder.

[0016] The target encoding result is input into the decoder of the target speech recognition network to obtain the speech recognition result of the target speech recognition network.

[0017] In this embodiment of the application, the feature vector of the speech waveform is input to a routing network so that the routing network selects at least one target speech recognition network corresponding to the feature vector from the plurality of speech recognition networks, including:

[0018] A multidimensional vector is determined based on the feature vector and a preset weight matrix; the row index i of the preset weight matrix represents the weight vector of the i-th speech recognition network, and the column index j represents the j-th component of the feature vector; the value of each dimension of the multidimensional vector is used to represent the score of the corresponding speech recognition network.

[0019] Based on the multidimensional vector, the top k speech recognition networks with the highest scores from the i speech recognition networks are selected as the target speech recognition network.

[0020] In this embodiment of the application, determining the final recognition result of the speech waveform based on the at least one speech recognition result includes:

[0021] The values ​​of multiple dimensions of the multidimensional vector are normalized to obtain at least one output weight of the at least one target speech recognition network.

[0022] The final recognition result of the speech waveform is obtained by weighted summation based on the at least one speech recognition result and the at least one output weight.

[0023] A second aspect of this application provides a speech recognition model, which includes a feature extraction network, multiple speech recognition networks, a routing network, and a result output network.

[0024] The feature extraction network is used to extract features from the acquired speech waveform to be recognized to obtain the feature vector of the speech waveform;

[0025] The routing network is used to select at least one target speech recognition network corresponding to the feature vector from the plurality of speech recognition networks;

[0026] Each target speech recognition network is used to perform speech recognition on the feature vector of the speech waveform to obtain the speech recognition result;

[0027] The result output network is used to determine the final recognition result of the speech waveform based on the at least one speech recognition result.

[0028] In this embodiment, the speech recognition model is an initial model trained using the CTC loss function, the load balancing loss function, and the dynamic K loss function. The CTC loss function characterizes the difference between the final recognition result output by the model and the reference result. The load balancing loss function characterizes the difference between the actual load distribution and the ideal uniform distribution of each speech recognition network on a batch of samples. The dynamic K loss function characterizes the difference between the actual number of speech recognition networks activated by the speech recognition model under the current sample complexity and the expected number of activations.

[0029] A third aspect of this application provides a speech recognition device, the device comprising:

[0030] The data acquisition module is used to acquire the speech waveform to be recognized;

[0031] A network filtering module is used to input the feature vector of the speech waveform into a routing network, so that the routing network filters out at least one target speech recognition network corresponding to the feature vector from multiple speech recognition networks;

[0032] A speech recognition result generation module is used to input the feature vectors of the speech waveform into the at least one target speech recognition network to obtain at least one speech recognition result;

[0033] The final recognition result determination module is used to determine the final recognition result of the speech waveform based on the at least one speech recognition result.

[0034] An embodiment of the fourth aspect of this application provides a computer device including a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect above.

[0035] An embodiment of the fifth aspect of this application provides a computer-readable storage medium storing computer instructions that are used to cause a computer to perform the method described in the first aspect above.

[0036] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0038] Figure 1 A schematic flowchart of a speech recognition method provided in an embodiment of this application is shown;

[0039] Figure 2 A flowchart illustrating another speech recognition method provided in an embodiment of this application is shown;

[0040] Figure 3 This invention provides a schematic diagram of the structure of a speech recognition device according to an embodiment of the present application.

[0041] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of this application is shown. Detailed Implementation

[0042] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0043] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0044] The technical scenarios involved in the embodiments of this application are described below.

[0045] With the development of artificial intelligence technology, speech recognition technology has been widely applied. Traditional automatic speech recognition (ASR) models typically employ end-to-end structures, such as models based on CTC, attention mechanisms, or Transformers. However, in multilingual or multi-task scenarios, these models often require large amounts of labeled data and are difficult to scale flexibly. In recent years, hybrid multi-expert networks (MoE) have been widely used to build large-scale language models. Their core idea is to selectively activate some experts through routing mechanisms, thereby achieving efficient model training and inference. CTC, as a classic time-series alignment method, is widely used in speech recognition tasks, especially suitable for handling problems where the input and output sequence lengths are inconsistent.

[0046] However, there is currently no complete method proposed to combine MoE and CTC for speech recognition, especially in multilingual, low-resource language recognition scenarios. There is a lack of a systematic solution that can utilize the expert selection capability of MoE and combine the sequence modeling advantages of CTC.

[0047] According to an embodiment of this application, a speech recognition method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0048] This embodiment provides a speech recognition method. Figure 1 This is a flowchart of a speech recognition method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0049] Step S101: Obtain the speech waveform to be recognized.

[0050] After acquiring the speech waveform to be recognized, the method further includes steps a1-a2:

[0051] Step a1: Extract features from the speech waveform to obtain the original acoustic features.

[0052] Step a2: Encode the original acoustic features into high-dimensional features with contextual semantic information, and use the high-dimensional features as the feature vector of the speech waveform.

[0053] In this embodiment, the speech waveform to be recognized is obtained through voice input, and the corresponding original acoustic features are obtained by feature extraction of the speech waveform to be recognized. Feature extraction can be performed by log-Mel filter bank or FBank, and the specific feature extraction method is not specifically limited. Then, the original acoustic features are feature encoded by a shared underlying encoder to obtain high-dimensional features with contextual semantic information as the feature vector of the speech waveform, which is then input into the routing network.

[0054] Step S102: Input the feature vector of the speech waveform into the routing network so that the routing network can select at least one target speech recognition network corresponding to the feature vector from multiple speech recognition networks.

[0055] In some specific embodiments, step S102 above includes steps S1021-S1022:

[0056] Step S1021: Determine the multidimensional vector based on the feature vector and the preset weight matrix.

[0057] Specifically, the row index i of the preset weight matrix represents the weight vector of the i-th speech recognition network, and the column index j represents the j-th component of the feature vector; the value of each dimension of the multidimensional vector is used to represent the score of the corresponding speech recognition network.

[0058] Step S1022: Select the top k speech recognition networks with the highest scores from the i speech recognition networks based on the multidimensional vector as the target speech recognition network.

[0059] Specifically, steps S1021-S1022 above will be explained using the following example:

[0060] Assume that the feature vector of the speech waveform is an 80-dimensional acoustic feature vector, i.e., v = [0.12, -0.35, 0.87, ..., 0.44] (a total of 80 components).

[0061] Assume there are 4 speech recognition subnetworks (A / B / C / D), with a weight matrix W ∈ (4*80), where each row corresponds to the weight vector of one subnetwork, for example:

[0062] W_A=[w_{A,1},w_{A,2},…,w_{A,80}]

[0063] W_B=[w_{B,1},w_{B,2},…,w_{B,80}]

[0064] …

[0065] These weights are fixed during the training phase.

[0066] Calculate the multidimensional score vector s = W·v (matrix multiplication) to obtain the 4-dimensional vector s = [2.10, -0.45, 3.24, 1.57], where s_A = 2.10, s_B = -0.45, s_C = 3.24, and s_D = 1.57.

[0067] Select the top k subnetworks, let k=2, and sort them in descending order of score: 1st place: C (3.24), 2nd place: A (2.10).

[0068] Therefore, subnetworks C and A are used as the "target speech recognition network" for this inference, while the remaining networks B and D are not included in the calculation.

[0069] Step S103: Input the feature vector of the speech waveform into the at least one target speech recognition network to obtain at least one speech recognition result.

[0070] Specifically, each speech recognition network includes an encoder and a decoder, for example... Figure 2 The “encoder 1” and “decoder 1” in the text belong to the same speech recognition network.

[0071] In some specific embodiments, step S103 above includes steps S1031-S1033:

[0072] Step S1031: For any target speech recognition network, determine a second sequence from the first sequence that corresponds to the target encoder of the target speech recognition network;

[0073] Specifically, the first sequence includes multiple encoders stacked serially in a preset order, for example... Figure 2 The sequence is “Encoder 1 → Encoder 2 → Encoder 3 → Encoder 4”. Each encoder corresponds one-to-one with a speech recognition network; the second sequence refers to the serial sequence in the first sequence that uses the target encoder as the tail encoder, for example… Figure 2 As shown: When the target encoder is encoder 3, the second sequence is "encoder 1 → encoder 2 → encoder 3".

[0074] Step S1032: Input the feature vector of the speech waveform into the second sequence to obtain the target encoding result of the target encoder.

[0075] Specifically, the encoding result output by the previous encoder in the second sequence is used as the input of the next encoder. For example, the encoding result output by encoder 1 can be used as the input of encoder 2.

[0076] Step S1033: Input the target encoding result into the decoder of the target speech recognition network to obtain the speech recognition result of the target speech recognition network.

[0077] Specifically, such as Figure 2 As shown, when the target speech recognition network includes encoder 2 and decoder 2, the feature vector of the speech waveform (a high-dimensional feature with contextual semantic information obtained by feature encoding) is input into encoder 1. The encoding result output by encoder 1 is then input into encoder 2, and the encoding result output by encoder 2 is then input into decoder 2, thereby obtaining the speech recognition result of the target speech recognition network. Other target speech recognition networks are similar and will not be elaborated here.

[0078] More specifically, some parameters of multiple speech recognition networks are shared; for example, the input of encoder 3 is the output of encoder 2.

[0079] Step S104: Determine the final recognition result of the speech waveform based on the at least one speech recognition result.

[0080] Specifically, the final recognition result can be obtained by weighted summation of each speech recognition result and its corresponding weight.

[0081] In some specific embodiments, step 104 above includes steps S1041-S1042:

[0082] Step S1041: Normalize the values ​​of multiple dimensions of the multidimensional vector to obtain at least one output weight of the at least one target speech recognition network.

[0083] Specifically, the weights are obtained by transforming the values ​​of multiple dimensions of a multidimensional vector using Softmax, for example:

[0084] Original score s = [2.10, -0.45, 3.24, 1.57]; After Softmax → weight w = [0.20, 0.05, 0.55, 0.20]

[0085] Step S1042: The final recognition result of the speech waveform is obtained by weighted summation based on the at least one speech recognition result and the at least one output weight.

[0086] Corresponding to the above implementation methods of speech recognition, this application also provides a speech recognition model:

[0087] The speech recognition model includes a feature extraction network, multiple speech recognition networks, a routing network, and a result output network.

[0088] The feature extraction network is used to extract features from the acquired speech waveform to be recognized to obtain the feature vector of the speech waveform;

[0089] The routing network is used to select at least one target speech recognition network corresponding to the feature vector from the plurality of speech recognition networks;

[0090] Each target speech recognition network is used to perform speech recognition on the feature vector of the speech waveform to obtain the speech recognition result;

[0091] The result output network is used to determine the final recognition result of the speech waveform based on the at least one speech recognition result.

[0092] In some specific embodiments, the above-mentioned speech recognition model is obtained by training the initial model using the CTC loss function, the load balancing loss function, and the dynamic K loss function.

[0093] Specifically, the CTC loss function is used to characterize the difference between the final recognition result output by the model and the reference result; the load balancing loss function is used to characterize the difference between the actual load distribution and the ideal uniform distribution of each speech recognition network on a batch of samples.

[0094] More specifically, the dynamic K-loss function is used to characterize the difference between the number of speech recognition networks actually activated by the speech recognition model under the current sample complexity and the expected number of activations. This difference is quantified by calculating the absolute value or squared error between the number of speech recognition networks assigned to the current sample or sequence and a preset learnable threshold, and scaled by a preset weight coefficient to encourage the model to reduce the number of activated speech recognition networks in simple tasks and increase the number of activated speech recognition networks in complex tasks, thereby reducing computational overhead while ensuring recognition accuracy.

[0095] Corresponding to the above implementation methods for speech recognition, this application also provides a speech recognition device for executing the speech recognition methods described in the above embodiments. For example... Figure 3 As shown, the voice recognition device includes:

[0096] The data acquisition module 301 is used to acquire the speech waveform to be recognized;

[0097] The network filtering module 302 is used to input the feature vector of the speech waveform into the routing network, so that the routing network filters out at least one target speech recognition network corresponding to the feature vector from multiple speech recognition networks;

[0098] The speech recognition result generation module 303 is used to input the feature vector of the speech waveform into the at least one target speech recognition network to obtain at least one speech recognition result;

[0099] The final recognition result determination module 304 is used to determine the final recognition result of the speech waveform based on the at least one speech recognition result.

[0100] Optionally, the apparatus further includes: a feature encoding module, configured to, after acquiring the speech waveform to be recognized, extract features from the speech waveform to obtain original acoustic features; encode the original acoustic features into high-dimensional features with contextual semantic information, and use the high-dimensional features as the feature vector of the speech waveform.

[0101] Optionally, the speech recognition result generation module is further configured to, for any target speech recognition network, determine a second sequence from a first sequence corresponding to the target encoder of the target speech recognition network; the first sequence includes multiple encoders stacked in a preset order, each of which corresponds one-to-one with the multiple speech recognition networks; the second sequence refers to the serial sequence in the first sequence with the target encoder as the tail encoder; the feature vector of the speech waveform is input into the second sequence to obtain the target encoding result of the target encoder; the encoding result output by the previous encoder in the second sequence is used as the input of the next encoder; the target encoding result is input into the decoder of the target speech recognition network to obtain the speech recognition result of the target speech recognition network.

[0102] Optionally, the network filtering module is further configured to determine a multidimensional vector based on the feature vector and a preset weight matrix; the row index i of the preset weight matrix represents the weight vector of the i-th speech recognition network, and the column index j represents the j-th component of the feature vector; the value of each dimension of the multidimensional vector is used to represent the score of the corresponding speech recognition network; and the top k speech recognition networks with the highest scores are selected from the i speech recognition networks based on the multidimensional vector as the target speech recognition network.

[0103] Optionally, the final recognition result determination module is further configured to normalize the values ​​of multiple dimensions of the multidimensional vector to obtain at least one output weight of the at least one target speech recognition network; and to obtain the final recognition result of the speech waveform by weighted summation based on the at least one speech recognition result and the at least one output weight.

[0104] The speech recognition device and the speech recognition method provided in the above embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods used, run or implemented by the applications stored in them.

[0105] This invention also provides a computer device having the above-described features. Figure 3 The voice recognition device shown.

[0106] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 4 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0107] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0108] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0109] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0110] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0111] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0112] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0113] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0114] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A voice recognition method, characterized by, The method comprises: acquiring a speech waveform to be recognized; inputting a feature vector of the speech waveform into a routing network to enable the routing network to select at least one target speech recognition network corresponding to the feature vector from a plurality of speech recognition networks; inputting the feature vector of the speech waveform into the at least one target speech recognition network respectively to obtain at least one speech recognition result; determining a final recognition result of the speech waveform according to the at least one speech recognition result.

2. The method of claim 1, wherein, After acquiring the speech waveform to be recognized, the method further comprises: performing feature extraction on the speech waveform to obtain original acoustic features; encoding the original acoustic features into high-dimensional features with context semantic information, and taking the high-dimensional features as the feature vector of the speech waveform.

3. The method according to claim 1 or 2, characterized in that, Each speech recognition network comprises an encoder and a decoder; inputting the feature vector of the speech waveform into the at least one target speech recognition network respectively to obtain at least one speech recognition result comprises: for any target speech recognition network, determining a second sequence corresponding to a target encoder of the target speech recognition network from a first sequence; the first sequence comprises a plurality of encoders stacked in series according to a preset order, and the plurality of encoders correspond to the plurality of speech recognition networks one by one; the second sequence refers to a serial sequence in the first sequence with the target encoder as a tailing encoder; inputting the feature vector of the speech waveform into the second sequence to obtain a target encoding result of the target encoder; an encoding result output by a previous encoder in the second sequence is taken as an input of a next encoder; inputting the target encoding result into a decoder of the target speech recognition network to obtain a speech recognition result of the target speech recognition network.

4. The method of claim 1, wherein, Inputting the feature vector of the speech waveform into the routing network to enable the routing network to select at least one target speech recognition network corresponding to the feature vector from the plurality of speech recognition networks comprises: determining a multi-dimensional vector according to the feature vector and a preset weight matrix; a row index i of the preset weight matrix represents a weight vector of an i th speech recognition network, and a column index j represents a j th component of the feature vector; a value of each dimension of the multi-dimensional vector is used to represent a score of a corresponding speech recognition network; selecting the top k speech recognition networks with the highest scores from the i speech recognition networks as the target speech recognition networks according to the multi-dimensional vector.

5. The method of claim 4, wherein, Determining the final recognition result of the speech waveform according to the at least one speech recognition result comprises: performing normalization processing on the values of the plurality of dimensions of the multi-dimensional vector to obtain at least one output weight of the at least one target speech recognition network; performing weighted summation according to the at least one speech recognition result and the at least one output weight to obtain the final recognition result of the speech waveform.

6. A speech recognition model, characterized in that, The speech recognition model comprises a feature extraction network, a plurality of speech recognition networks, a routing network and a result output network; the feature extraction network is configured to perform feature extraction on the acquired speech waveform to be recognized to obtain a feature vector of the speech waveform; The routing network is configured to select at least one target speech recognition network corresponding to the feature vector from the plurality of speech recognition networks; Each target speech recognition network is configured to perform speech recognition on the feature vector of the speech waveform to obtain a speech recognition result. The result output network is configured to determine a final recognition result of the speech waveform according to the at least one speech recognition result.

7. The speech recognition model of claim 6, wherein, The speech recognition model is obtained by training an initial model using a CTC loss function, a load balancing loss function, and a dynamic K loss function. The CTC loss function is used to represent the difference between the final recognition result output by the model and a reference result. The load balancing loss function is used to represent the difference between the actual load distribution of each speech recognition network on a batch of samples and an ideal uniform distribution. The dynamic K loss function is used to represent the difference between the actual number of activated speech recognition networks of the speech recognition model at the current sample complexity and the expected number of activated speech recognition networks.

8. A speech recognition apparatus characterized by comprising: The apparatus comprises: a data acquisition module configured to acquire a speech waveform to be recognized; a network screening module configured to input a feature vector of the speech waveform into a routing network, so that the routing network selects at least one target speech recognition network corresponding to the feature vector from a plurality of speech recognition networks; a speech recognition result generation module configured to input the feature vector of the speech waveform into the at least one target speech recognition network respectively, to obtain at least one speech recognition result; a final recognition result determination module configured to determine a final recognition result of the speech waveform according to the at least one speech recognition result.

9. A computer device, comprising: comprise: a memory and a processor, which are communicatively connected, and the memory stores computer instructions, and the processor executes the computer instructions to perform the method of any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to perform the method of any one of claims 1 to 5.

Images