Two-stage event-driven wake-up circuit and SoC chip
Through a two-stage event-driven wake-up circuit, the voice activity detection of normally open event detection and in-memory computing is used to distinguish events from noise, reduce false wake-up and power consumption waste, realize efficient event processing, and solve the error in the existing technology Energy efficiency problems caused by awakening.
Patent Information
- Application Number
- CN202410772184.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-16
- Publication Date
- 2024-11-08
AI Technical Summary
In the prior art, event-driven wake-up technology has a problem of false wake-up, which leads to waste of power consumption, especially in noise or unnecessary situations, resulting in invalid wake-up, resulting in waste of effective computing power and energy efficiency.
Two-level event-driven wake-up circuit is adopted. The first level is the normally open event activity detection, and the second level performs further detection according to the specific application configuration. If a specific activity event is detected, the subsequent event processing part will be awakened, and the event activity detector and voice are used. The activity detector distinguishes event sound from background noise, and identifies the vocal signal in in-memory calculations, and finally wakes up the high-power DNN accelerator and Cortex-M4 CPU core.
It effectively reduces the false wake-up rate and long-term average power consumption, reduces energy consumption by an additional 87.3%, and improves the energy efficiency of the system.
Smart Images

Figure CN118915894A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of integrated circuits, and in particular to a two-level event-driven wake-up circuit and a SoC chip. Background Art
[0002] Random and sparse application scenarios often have a large amount of invalid / redundant data. To detect the occurrence of events, the traditional approach is usually a periodic polling mechanism. To avoid missing events, in actual applications, the frequency of "sleep-wake-up" is much higher than the actual occurrence rate of events, resulting in a waste of effective computing power / energy efficiency. The event-driven wake-up technology that uses an always-on ultra-low power circuit with certain event detection capabilities to wake up a high-performance, high-power circuit can avoid the power waste caused by periodic polling.
[0003] In the prior art, event-driven wake-up technology is usually directly based on the occurrence of analog signals such as voltage amplitude or signal frequency to detect the occurrence of events. When an event is detected, it drives the power-consuming CPU or other complex computing modules to turn on. However, when noise or other unnecessary false wake-ups occur, power consumption will still be wasted. Summary of the invention
[0004] The present invention provides a two-level event-driven wake-up circuit and a SoC chip, which can effectively reduce false wake-ups and further reduce power consumption.
[0005] The present invention provides a two-stage event-driven wake-up circuit, comprising:
[0006] The first level event detection part is used for normally open event activity detection;
[0007] The second level event detection part performs event detection on the first level detection results according to the specific activity detection configured by the specific application;
[0008] When the first stage detects an activity event, the second stage event detection part is awakened to start working, so as to further perform event detection on the first stage detection result. If a specific activity event is detected, the event processing part of the subsequent stage is awakened.
[0009] The present invention also provides a SoC chip, comprising an event processing part and the above-mentioned two-level event driven wake-up circuit. When the two-level event driven wake-up circuit detects a specific activity event, the event processing part starts working.
[0010] The present invention adopts a two-level event-driven chip wake-up technology based on in-memory calculation, which can avoid the problem of high power consumption caused by false wake-up caused by the first-level wake-up. Through the two-level wake-up, for voice keyword applications, the two-level event wake-up uses an event activity detector (EAD) to distinguish event sounds from background noise, and wakes up the subsequent voice activity detection (VAD) and digital signal processing circuits; the voice activity detection (VAD) based on in-memory calculation identifies whether the signal is a human voice, and then wakes up the subsequent DNN accelerator and Cortex-M4 CPU core with higher power consumption, which can effectively minimize the long-term average power consumption and false wake-up rate, and can additionally reduce energy consumption by 87.3%. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0012] Figure 1 A schematic diagram of a two-level event-driven wake-up circuit provided by an embodiment of the present invention;
[0013] Figure 2 This is a schematic diagram of the workflow of two-level event-driven wakeup;
[0014] Figure 3 This is the timing diagram for two-level event-driven wake-up;
[0015] Figure 4 Schematic diagram of the circuit structure of the first level event activity detection in an embodiment of the present invention;
[0016] Figure 5 is a schematic diagram of a second-level voice activity detector module based on in-memory computing;
[0017] Figure 6 A comparative schematic diagram of energy consumption savings when an embodiment of the present invention is applied in a KWS scenario;
[0018] Figure 7 Schematic diagram of the edge keyword recognition SoC architecture and end-to-end edge AI processing flow provided by the present invention. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0020] During the end-to-end edge machine learning TinyML processing, the SoC is in standby mode most of the time, waiting for valid input audio signals. The wake-up technology in the prior art usually adopts a one-level wake-up, that is, directly judging whether to wake up the high-energy consumption subsequent module based on the voltage amplitude or signal frequency. In this case, there is generally a false wake-up or an unnecessary wake-up. Therefore, the present invention provides a two-level event-driven wake-up technology to solve these problems and further reduce power consumption.
[0021] To make the technical solution of the present invention clearer, embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0022] Figure 1 A schematic diagram of a two-stage event-driven wake-up circuit provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the two-level event-driven wake-up circuit in this embodiment includes: a first-level event detection part 10, which is used for normally open event activity detection; a second-level event detection part 20, which performs event detection on the first-level detection result according to the specific activity detection requirements of the specific application configuration; when the first level detects an activity event, the second-level event detection part 20 is awakened to start working, so as to further perform event detection on the first-level detection result, and if a specific activity event is detected, the event processing part 30 of the subsequent stage is awakened.
[0023] In specific applications, such as Figure 2 The workflow diagram of the two-level event-driven wake-up is shown. If the event processing part of the latter stage is a keyword recognition and reasoning module containing a neural network accelerator, the first-level event detection part includes an event activity detection module, which is used to detect and identify sound information, that is, a sound collection device, such as a microphone, collects sound information in the environment, and the first-level event detection part is configured with a normally open event activity detection; the second-level event detection part includes a voice activity detection module and a digital signal processing module, and the voice activity detection module identifies the sound information detected by the event activity detection module in the first stage to determine whether the sound information is human voice information; when human voice information is detected, the keyword recognition and reasoning module identifies and analyzes the human voice information detected by the voice activity detection module to determine whether it contains preset keywords, so as to drive other execution modules to perform corresponding processing. Figure 2The specific working process of the two-level wake-up circuit module is as follows Figure 3 The timing diagram shown in the figure shows the waveform of the workflow. The first level has only normally open modules, mainly event activity detector and cortex-m0CPU working, the second level has only digital signal processing and voice activity detector working, and the third level is keyword recognition reasoning, cortex-m4CPU, digital signal processing and deep neural network accelerator are all working. The blue arrow indicates wake-up, and it can be seen that the power consumption will gradually increase with event wake-up. The state change of the frame number indicates that the input signal is processed by DSP. After calculating 6 frames of data, 6 frames of data are sent to the VAD module for voice recognition; after calculating 60 frames of data, 60 frames of data are sent to the DNN module for keyword recognition reasoning.
[0024] In the above embodiment, the event activity detection module includes an event starting point detection part and an end point detection part. Both the event starting point detection part and the end point detection part use parallel amplitude threshold detection and zero-crossing rate detection on the collected sound information, and use the earlier time in the detection result as the event starting point. Figure 4 The circuit structure diagram of the first-level event activity detection in the embodiment of the present invention is shown in FIG. The first-level EAD module determines the start and end of the sound according to the voice signal swing and zero-crossing rate, and outputs a valid signal to wake up the second-level VAD based on in-memory calculation (CIM). First, for the starting point detection, the original voice signal output by the microphone is subjected to amplitude threshold detection and zero-crossing rate detection in parallel. The one whose detection result is earlier in time is used as the starting point. For the zero-crossing rate detection, an offset needs to be added to make the zero-crossing rate of pure noise 0. The starting point triggers the "volume calculation" module to work and convert the original voice waveform into a volume waveform. If the amplitude threshold judgment result is earlier, in order not to miss the event, the volume judgment will be used as an auxiliary, because the operation of volume calculation can expand the difference between noise and events; for the end point detection, the original voice signal output by the microphone is also subjected to amplitude threshold detection and zero-crossing rate detection in parallel, and a configurable time window is additionally observed to avoid misjudgment of the end point due to "polysyllabic words" or "short intervals of words within a phrase".
[0025] Figure 5 FIG. 1 is a schematic diagram of a second-level voice activity detector module based on in-memory computing. Figure 5 As shown, Figure 5The upper part shows the CNN model to be deployed for voice activity detection. At the algorithm level, convolution 1 layer (conv1), convolution 2 layers (conv2), maximum pooling layer (max_pooling), fully connected 1 layer (FC2), and fully connected 2 layers (FC2) are sequentially performed. Correspondingly, at the architecture level, for smaller Conv1 and FC2, RTL is directly used for implementation; for larger CONV2 and FC1, unified CIM IP is used to reduce the workload of the backend. The architecture of the digital domain in-memory computing array is shown in the figure, including a DCIM array based on 6T SRAM, a word line decoder, an input buffer, and a global read and write circuit. The DCIM array consists of 16 DCIM columns. Each DCIM column consists of 16 rows of sub-arrays and an addition tree for performing vector-to-vector multiplication and accumulation (MAC) operations. Each sub-array contains 16 rows and 8 columns of 6T SRAM cells, N local readout and calculation modules, and 2N configuration switches. Each row stores an 8-bit weight, and each sub-array stores a total of 16 weights. The addition tree sums the product results and sends the sum to the shift accumulator to achieve the expansion of input precision in the time dimension. A CNN model with an accuracy of more than 90% is fully deployed in the on-chip digital CIM macro, which can process 16×1-bit input and 8-bit weights per cycle. CIM adopts a 256×128 array design with an efficiency of 4.4TOPS / W at INT8 precision.
[0026] The embodiment of the present invention adopts a two-level event-driven chip wake-up technology based on in-memory calculation, which can avoid the problem of high power consumption caused by false wake-up caused by the first-level wake-up. Through the two-level wake-up, for voice keyword applications, the two-level event wake-up uses an event activity detector (EAD) to distinguish event sounds from background noise, and wakes up the subsequent voice activity detection (VAD) and digital signal processing circuits; the voice activity detection (VAD) based on in-memory calculation identifies whether the signal is a human voice, and then wakes up the subsequent higher power consumption DNN accelerator and Cortex-M4 CPU core, which can effectively minimize the long-term average power consumption and false wake-up rate, and can additionally reduce energy consumption by 87.3%.
[0027] Figure 6 FIG. 1 is a comparative schematic diagram of energy saving in a KWS scenario when an embodiment of the present invention is applied. Figure 6 As shown, in a typical KWS scenario, such as a sound activation rate of 33.3% (8 hours / day) and a voice activation rate of 8.3% (2 hours / day), compared with a design without event-driven, a single-stage event-driven design can save 71.6% of energy, while the two-stage wake-up of the present invention can reduce energy consumption by an additional 87.3%.
[0028] An embodiment of the present invention further provides a SoC chip, comprising an event processing part and the above-mentioned two-level event driven wake-up circuit. When the two-level event driven wake-up circuit detects a specific activity event, the event processing part starts working. Figure 7 The edge keyword recognition SoC architecture and end-to-end edge AI processing flow diagram provided by the present invention are as follows: Figure 7 As shown, the SoC chip architecture includes an always-on Cortex-M0 CPU, a deep neural network (DNN) accelerator for edge keyword detection reasoning, a self-designed digital signal processing circuit (digital signal processing, DSP) for audio signal processing, and a Cortex-M4 CPU for acoustic echo cancellation (AEC). In addition, the architecture adopts an on-chip global storage system based on SRAM for sharing data between different modules. For the workflow of two-level event-driven wakeup: the present invention adopts a two-level event-driven wakeup, in which the first level is distinguished from the background noise by the always-on (normally open) event activity detector (event activity detection, EAD), and wakes up the digital signal processing circuit and voice activity detector (voice activity detection, VAD) of the subsequent stage; the second level is based on the computing in memory (computing in memory, CIM) voice activity detector (voice activity detection, VAD) to identify whether the signal is a human voice, and then wake up the neural network accelerator and Cortex-M4 CPU core with higher power consumption of the subsequent stage. This pipelined two-stage wake-up effectively minimizes long-term average power consumption and false wake-up rate. The chip also uses a centralized power management module to monitor the runtime energy, architectural events, and performance of all heterogeneous blocks, and adjusts the voltage and frequency of each block to achieve optimal system energy efficiency. The chip operates in different modes. During end-to-end edge AI processing, the SoC remains in standby mode most of the time, with only the EAD and Cortex-M0 active. The audio signal is processed by the EAD to determine the beginning and end of the sound. Next is the VAD mode, where the audio signal is sent to the DSP and the CIM-based VAD module to distinguish between event sounds and background noise. Finally, in the KWS inference mode, the audio signal is processed by the Cortex-M4 for acoustic echo cancellation, the DSP for audio signals, and the MEM system and DNN accelerator for KWS inference.
[0029] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A two-level event-driven wake-up circuit, It is characterized in that include: The first level event detection part is used for normally open event activity detection; The second level event detection part performs event detection on the first level detection results according to the specific activity detection configured by the specific application; When the first stage detects an activity event, the second stage event detection part is awakened to start working, so as to further perform event detection on the first stage detection result. If a specific activity event is detected, the event processing part of the subsequent stage is awakened.
2. The circuit according to claim 1, It is characterized in that When the event processing part of the latter stage is a keyword recognition reasoning module containing a neural network accelerator, the first-level event detection part includes an event activity detection module for detecting and identifying sound information; the second-level event detection part includes a voice activity detection module and a digital signal processing module, and the voice activity detection module identifies the sound information detected by the event activity detection module to determine whether the sound information is human voice information; The keyword recognition and reasoning module recognizes and analyzes the human voice information detected by the voice activity detection module to determine whether it contains preset keywords.
3. The circuit according to claim 2, It is characterized in that The event activity detection module includes an event starting point detection part and an end point detection part. Both the event starting point detection part and the end point detection part use parallel amplitude threshold detection and zero-crossing rate detection on the collected sound information, and use the earlier time in the detection result as the event starting point.
4. The circuit according to claim 3, It is characterized in that Applies a noise-removing offset to the zero-crossing rate detection.
5. The circuit according to claim 3, It is characterized in that If the amplitude threshold detection result is at the front, volume judgment is used as an auxiliary.
6. The circuit according to claim 3, It is characterized in that The end point detection part is configured with a time window for additional observation.
7. The circuit according to any one of claims 1 to 6, It is characterized in that The second level event detection part is the specific activity detection based on in-memory calculation.
8. A SoC chip, It is characterized in that The invention comprises an event processing part and a two-stage event driven wake-up circuit as claimed in any one of claims 1 to 7. When the two-stage event driven wake-up circuit detects a specific activity event, the event processing part starts to work.
Citation Information
Cited By
Low-power-consumption voice recognition and local task buffering method for AI glasses
CN121789655A
Low-power voice recognition and local task buffering for ai glasses
CN121789655B
