This document describes the process of setting up and running an example of sequential execution of the Keyword Spotting and Automatic Speech Recognition models on a Cortex-M CPU and Ethos-U NPU.
The Keyword Spotting and Automatic Speech Recognition example demonstrates how to run multiple models sequentially.
A Keyword Spotting model is first run on the CPU. If a set keyword is detected on the remaining audio, then an Automatic Speech Recognition model is run on the Ethos-U NPU.
The tensor arena memory region is reused between models to optimize application memory footprint.
The Yes keyword is used to trigger full command recognition following the keyword.
Use-case code could be found in the following directory: source/use_case/kws_asr.
In this use-case, there are two different models being used with different requirements for preprocessing. As such, each preprocessing process is detailed as follows.
Note: Automatic Speech Recognition only occurs if a keyword is detected in the audio clip.
By default, the KWS model is run purely on the CPU and not on the Ethos-U55.
The MicroNet keyword spotting model that is used with the Code Samples expects audio data to be preprocessed in a specific way before performing an inference.
Therefore, this section aims to provide an overview of the feature extraction process used.
First, the audio data is normalized to the range (-1, 1).
Note: Mel-Frequency Cepstral Coefficients (MFCCs) are a common feature that is extracted from audio data and can be used as input for machine learning tasks such as keyword spotting and speech recognition. For implementation details, please refer to:
source/application/main/include/Mfcc.hpp
Next, a window of 640 audio samples is taken from the start of the audio clip. From these 640 samples, we calculate 10 MFCC features.
The whole window is shifted to the right by 320 audio samples and 10 new MFCC features are calculated. This process of shifting and calculating is repeated until the end of the 16000 audio samples required to perform an inference is reached.
In total, this is 49 windows that each have 10 MFCC features calculated for them, giving an input tensor of shape 49x10.
These extracted features are quantized and an inference is performed.
If the audio clip is longer than 16000 audio samples, then the initial starting position is offset by 16000/2 = 8000 audio samples. From this new starting point, MFCC features for the next 16000 audio samples are calculated and another inference is performed. In other words, do an inference for samples 8000-24000.
Note: Parameters of the MFCC feature extraction all depend on what was used during model training. These values are specific to each model.
If you try a different keyword spotting model that uses MFCC input, then values check to see if the values need changing to match the new model.
In addition, MFCC feature extraction methods can vary slightly with different normalization methods or scaling being used.
The wav2letter automatic speech recognition model that is used with the Code Samples expects audio data to be preprocessed in a specific way before performing an inference. This section aims to provide an overview of the feature extraction process used.
First, the audio data is normalized to the range (-1, 1).
Note: Mel-Frequency Cepstral Coefficients (MFCCs) are a common feature that is extracted from audio data and can be used as input for machine learning tasks. Such as keyword spotting and speech recognition. For implementation details, please refer to:
source/application/main/include/Mfcc.hpp
Next, a window of 512 audio samples is taken from the start of the audio clip. From these 512 samples, we calculate 13 MFCC features.
The whole window is shifted to the right by 160 audio samples and 13 new MFCC features are calculated. This process of shifting and calculating is repeated until enough audio samples to perform an inference have been processed.
In total, this is 296 windows that each have 13 MFCC features calculated for them.
After extracting MFCC features, the first and second order derivatives of these features, regarding time, are calculated.
These derivative features are then standardized and concatenated with the MFCC features (which also get standardized). At this point, the input tensor has a shape of 296x39.
These extracted features are quantized and an inference is performed.
For longer audio clips, where multiple inferences must be performed, then the initial starting position is offset by (100*160) = 16000 audio samples. From this new starting point, MFCC and derivative features are calculated as before, until there is enough to perform another inference.
Padding can be used if there are not enough audio samples for at least one inference. This step is repeated until the whole audio clip has been processed. If there are not enough audio samples for a final complete inference, then the MFCC features are padded by repeating the last calculated feature until an inference can be performed.
Note: Parameters of the MFCC feature extraction all depend on what was used during model training. These values are specific to each model.
If you switch to a different ASR model than the one supplied, then the feature extraction process could be completely different to the one currently implemented.
The amount of time that audio samples that are offset for long audio clips is specific to the included wav2letter model.
If a keyword is detected, then the ASR process is run and the raw output of that inference must be postprocessed to get a usable result.
The raw output from the model is a tensor of shape 148x29 where each row is a probability distribution over the possible 29 characters that can appear at each of the 148 time steps.
This wav2letter model is trained using context windows. This means that, depending on the bit of the audio clip that is being processed, only certain parts of the output are usable.
If this is the first inference, and multiple inferences are required, then ignore the final 49 rows of the output. Similarly, if this is the final inference from multiple inferences, then ignore the first 49 rows of the output.
Finally, if this inference is not the last, or the first inference, then ignore the first and last 49 rows of the model output.
Note: If the audio clip is small enough, then the whole of the model output is usable and there is no need to throw away any of the outputs before continuing.
Once any rows have been removed, the final processing can be done. To process the output, the letter with the highest probability at each time step is found first. Next, any letters that are repeated multiple times in a row are removed.
For example: [t, t, t, o, p, p] becomes [t, o, p]). Finally, the 29th blank token letter is removed from the output.
For the final output, the results from all inferences are combined before decoding. What you are left with is then displayed to the console.
See Prerequisites
In addition to the already specified build option in the main documentation, the Keyword Spotting and Automatic Speech Recognition use-case adds:
kws_asr_MODEL_TFLITE_PATH_ASR and kws_asr_MODEL_TFLITE_PATH_KWS: The path to the NN model file in TFLite format. The model is processed and then included into the application axf file. The default value points to one of the delivered set of models. Note that the parameters kws_asr_LABELS_TXT_FILE_KWS, kws_asr_LABELS_TXT_FILE_ASR,TARGET_PLATFORM, and ETHOS_U_NPU_ENABLED must be aligned with the chosen model. In other words:
ETHOS_U_NPU_ENABLED is set to On or 1, then the NN model is assumed to be optimized. The model naturally falls back to the Arm® Cortex®-M CPU if an unoptimized model is supplied.ETHOS_U_NPU_ENABLED is set to Off or 0, then the NN model is assumed to be unoptimized. Supplying an optimized model in this case results in a runtime error.kws_asr_FILE_PATH: The path to the directory containing audio files, or a path to single WAV file, to be used in the application. The default value points to the resources/kws_asr/samples folder that contains the delivered set of audio clips.
kws_asr_LABELS_TXT_FILE_KWS and kws_asr_LABELS_TXT_FILE_ASR: The respective paths to the keyword spotting labels and the automatic speech recognition labels text files. The file is used to map letter class index to the text label. The default value points to the delivered labels.txt file inside the delivery package.
kws_asr_AUDIO_RATE: The input data sampling rate. Each audio file from kws_asr_FILE_PATH is preprocessed during the build to match the NN model input requirements. The default value is 16000.
kws_asr_AUDIO_MONO: If set to ON, then the audio data is converted to mono. The default value is ON.
kws_asr_AUDIO_OFFSET: Begins loading audio data and starts from this specified offset, defined in seconds. the default value is set to 0.
kws_asr_AUDIO_DURATION: The length of the audio data to be used in the application in seconds. The default is 0, meaning that the whole audio file is used.
kws_asr_AUDIO_MIN_SAMPLES: Minimum number of samples required by the network model. If the audio clip is shorter than this number, then it is padded with zeros. The default value is 16000.
kws_asr_MODEL_SCORE_THRESHOLD_KWS: Threshold value that must be applied to the inference results for a label to be deemed valid. The default is 0.9.
kws_asr_MODEL_SCORE_THRESHOLD_ASR: Threshold value that must be applied to the automatic speech recognition inference results for a label to be deemed valid. The default is 0.5.
kws_asr_ACTIVATION_BUF_SZ: The intermediate, or activation, buffer size reserved for the NN model. By default, it is set to 2MiB and is enough for most models.
To ONLY build the automatic speech recognition example application, add -DUSE_CASE_BUILD=kws_asr to the cmake command line, as specified in: Building.
Note: This section describes the process for configuring the build for the MPS3: SSE-300. To build for a different target platform, please refer to: Building.
To build only the keyword spotting and automatic speech recognition example, create a build directory and navigate inside, like so:
mkdir build_kws_asr && cd build_kws_asr
On Linux, when providing only the mandatory arguments for CMake configuration, execute the following command to build only the Keyword Spotting and Automatic Speech Recognition application to run on the Ethos-U55 Fast Model:
cmake ../ -DUSE_CASE_BUILD=kws_asr
To configure a build that can be debugged using Arm DS specify the build type as Debug and then use the Arm Compiler toolchain file:
cmake .. \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/toolchains/bare-metal-armclang.cmake \ -DCMAKE_BUILD_TYPE=Debug \ -DUSE_CASE_BUILD=kws_asr
For further information, please refer to:
Note: If re-building with changed parameters values, we recommend that you clean the build directory and re-run the CMake command.
If the CMake command succeeds, build the application as follows:
make -j4
To see compilation and link details, add VERBOSE=1.
Results of the build are placed under the build/bin folder, like so:
bin ├── ethos-u-kws_asr.axf ├── ethos-u-kws_asr.htm ├── ethos-u-kws_asr.map └── sectors ├── images.txt └── kws_asr ├── ddr.bin └── itcm.bin
The bin folder contains the following files:
ethos-u-kws_asr.axf: The built application binary for the Keyword Spotting and Automatic Speech Recognition use-case.
ethos-u-kws_asr.map: Information from building the application. For example: The libraries used, what was optimized, and the location of objects.
ethos-u-kws_asr.htm: Human readable file containing the call graph of application functions.
sectors/kws_asr: Folder containing the built application. It is split into files for loading into different FPGA memory regions.
sectors/images.txt: Tells the FPGA which memory regions to use for loading the binaries in the sectors/.. folder.
The application anomaly detection is set up to perform inferences on data found in the folder, or an individual file, that is pointed to by the parameter kws_asr_FILE_PATH.
To run the application with your own audio clips, first create a folder to hold them and then copy the custom clips into the following folder:
mkdir /tmp/custom_files cp custom_audio1.wav /tmp/custom_files/
Note: Clean the build directory before re-running the CMake command.
Next, when building, set kws_asr_FILE_PATH to the location of the following folder:
cmake .. \ -Dkws_asr_FILE_PATH=/tmp/custom_files/ \ -DUSE_CASE_BUILD=kws_asr
The audio flies found in the kws_asr_FILE_PATH folder are picked up and automatically converted to C++ files during the CMake configuration stage. They are then compiled into the application during the build phase for performing inference with.
The log from the configuration stage tells you what audio directory path has been used:
-- User option kws_asr_FILE_PATH is set to /tmp/custom_files
After compiling, your custom inputs have now replaced the default ones in the application.
The application performs KWS inference using the model pointed to by the CMake parameter kws_asr_MODEL_TFLITE_PATH_KWS. ASR inference is performed using the model pointed to by the CMake parameter kws_asr_MODEL_TFLITE_PATH_ASR.
This section assumes you want to change the existing ASR model to a custom one. If, instead, you want to change the KWS model, then the instructions are the same. Except ASR changes to KWS.
Note: If you want to run the model using an Ethos-U, ensure that your custom model has been successfully run through the Vela compiler before continuing.
For further information: Optimize model with Vela compiler.
To run the application with a custom model, you must provide a labels_<model_name>.txt file of labels that are associated with the model. Each line of the file must correspond to one of the outputs in your model. Refer to the provided labels_wav2letter.txt file for an example.
Then, you must set kws_asr_MODEL_TFLITE_PATH to the location of the Vela processed model file and kws_asr_LABELS_TXT_FILEto the location of the associated labels file.
For example:
cmake .. \ -Dkws_asr_MODEL_TFLITE_PATH_ASR=<path/to/custom_asr_model_after_vela.tflite> \ -Dkws_asr_LABELS_TXT_FILE_ASR=<path/to/labels_custom_model.txt> \ -DUSE_CASE_BUILD=kws_asr
Note: Clean the build directory before re-running the CMake command.
The .tflite model files pointed to by kws_asr_MODEL_TFLITE_PATH_KWS and kws_asr_MODEL_TFLITE_PATH_ASR, and the labels text files pointed to by kws_asr_LABELS_TXT_FILE_KWS and kws_asr_LABELS_TXT_FILE_ASR are converted to C++ files during the CMake configuration stage. They are then compiled into the application for performing inference with.
The log from the configuration stage tells you what model path and labels file have been used, for example:
-- User option TARGET_PLATFORM is set to mps3 -- User option kws_asr_MODEL_TFLITE_PATH_ASR is set to <path/to/custom_asr_model_after_vela.tflite> ... -- User option kws_asr_LABELS_TXT_FILE_ASR is set to <path/to/labels_custom_model.txt> ... -- Using <path/to/custom_asr_model_after_vela.tflite> ++ Converting custom_asr_model_after_vela.tflite to\ custom_asr_model_after_vela.tflite.cc -- Generating labels file from <path/to/labels_custom_model.txt> -- writing to Labels_wav2letter ...
After compiling, your custom model has now replaced the default one in the application.
The FVP is available publicly from Arm Ecosystem FVP downloads.
For the Ethos-U evaluation, please download the MPS3 based version of the Arm® Corstone™-300 model that contains Cortex-M55 and offers a choice of the Ethos-U55 and Ethos-U65 processors.
To install the FVP:
Unpack the archive.
Run the install script in the extracted package:
./FVP_Corstone_SSE-300.sh
Once the building has been completed, the application binary ethos-u-kws_asr.axf can be found in the build/bin folder.
Assuming that the install location of the FVP was set to ~/FVP_install_location, then the simulation can be started by using:
$ ~/FVP_install_location/models/Linux64_GCC-6.4/FVP_Corstone_SSE-300_Ethos-U55 ./bin/mps3-sse-300/ethos-u-kws_asr.axf
A log output appears on the terminal:
telnetterminal0: Listening for serial connection on port 5000 telnetterminal1: Listening for serial connection on port 5001 telnetterminal2: Listening for serial connection on port 5002 telnetterminal5: Listening for serial connection on port 5003
This also launches a telnet window with the standard output of the sample application. It also includes error log entries containing information about the pre-built application version, TensorFlow Lite Micro library version used, and data types. The log also includes the input and output tensor sizes of the model compiled into the executable binary.
After the application has started, if kws_asr_FILE_PATH points to a single file, or even a folder that contains a single input file, then the inference starts immediately. If there are multiple inputs, it outputs a menu and then waits for input from the user.
For example:
User input required Enter option number from: 1. Classify next audio clip 2. Classify audio clip at chosen index 3. Run classification on all audio clips 4. Show NN model info 5. List audio clips Choice:
What the preceding choices do:
Classify next audio clip: Runs a single inference on the next in line.
Classify audio clip at chosen index: Runs inference on the chosen audio clip.
Note: Please make sure to select audio clip index within the range of supplied audio clips during application build. By default, a pre-built application has four files, with indexes from
0to3.
Run ... on all: Triggers sequential inference executions on all built-in applications.
Show NN model info: Prints information about the model data type, input, and output, tensor sizes:
INFO - Model info: INFO - Model INPUT tensors: INFO - tensor type is INT8 INFO - tensor occupies 490 bytes with dimensions INFO - 0: 1 INFO - 1: 49 INFO - 2: 10 INFO - 3: 1 INFO - Quant dimension: 0 INFO - Scale[0] = 0.201095 INFO - ZeroPoint[0] = -5 INFO - Model OUTPUT tensors: INFO - tensor type is INT8 INFO - tensor occupies 12 bytes with dimensions INFO - 0: 1 INFO - 1: 12 INFO - Quant dimension: 0 INFO - Scale[0] = 0.056054 INFO - ZeroPoint[0] = -54 INFO - Activation buffer (a.k.a tensor arena) size used: 127068 INFO - Number of operators: 1 INFO - Operator 0: ethos-u INFO - Model INPUT tensors: INFO - tensor type is INT8 INFO - tensor occupies 11544 bytes with dimensions INFO - 0: 1 INFO - 1: 296 INFO - 2: 39 INFO - Quant dimension: 0 INFO - Scale[0] = 0.110316 INFO - ZeroPoint[0] = -11 INFO - Model OUTPUT tensors: INFO - tensor type is INT8 INFO - tensor occupies 4292 bytes with dimensions INFO - 0: 1 INFO - 1: 1 INFO - 2: 148 INFO - 3: 29 INFO - Quant dimension: 0 INFO - Scale[0] = 0.003906 INFO - ZeroPoint[0] = -128 INFO - Activation buffer (a.k.a tensor arena) size used: 4184332 INFO - Number of operators: 1 INFO - Operator 0: ethos-u
List audio clips: Prints a list of pair ... indexes. The original filenames are embedded in the application, like so:
[INFO] List of Files: [INFO] 0 => yes_no_go_stop.wav
Please select the first menu option to execute Keyword Spotting and Automatic Speech Recognition.
The following example illustrates the output of an application:
INFO - KWS audio data window size 16000 INFO - Running KWS inference on audio clip 0 => yes_no_go_stop.wav INFO - Inference 1/7 INFO - For timestamp: 0.000000 (inference #: 0); threshold: 0.900000 INFO - label @ 0: yes, score: 0.997407 INFO - Profile for Inference: INFO - NPU AXI0_RD_DATA_BEAT_RECEIVED beats: 132130 INFO - NPU AXI0_WR_DATA_BEAT_WRITTEN beats: 48252 INFO - NPU AXI1_RD_DATA_BEAT_RECEIVED beats: 17544 INFO - NPU ACTIVE cycles: 413814 INFO - NPU IDLE cycles: 358 INFO - NPU TOTAL cycles: 414172 INFO - Keyword spotted INFO - Inference 1/2 INFO - Inference 2/2 INFO - Result for inf 0: no go INFO - Result for inf 1: stop INFO - Final result: no go stop INFO - Profile for Inference: INFO - NPU AXI0_RD_DATA_BEAT_RECEIVED beats: 8895431 INFO - NPU AXI0_WR_DATA_BEAT_WRITTEN beats: 1890168 INFO - NPU AXI1_RD_DATA_BEAT_RECEIVED beats: 1740069 INFO - NPU ACTIVE cycles: 30164330 INFO - NPU IDLE cycles: 342 INFO - NPU TOTAL cycles: 30164672 INFO - Main loop terminated. INFO - program terminating... INFO - releasing platform Arm Corstone-300 (SSE-300)
It can take several minutes to complete one inference run. The average time is around 2-3 minutes.
Using the input yes_no_go_stop.wav, the log shows inference results for the KWS operation first. Detecting the trigger word yes with the stated probability score. In this case, 0.99. After this, the ASR inference is run, printing the words recognized from the input sample.
The profiling section of the log shows that for the ASR inference:
Ethos-U PMU report:
30,164,672 total cycle: The number of NPU cycles.
30,164,330 active cycles: The number of NPU cycles that were used for computation.
342 idle cycles: The number of cycles for which the NPU was idle.
8,895,431 AXI0 read beats: The number of AXI beats with read transactions from the AXI0 bus. AXI0 is the bus where the Ethos-U NPU reads and writes to the computation buffers, activation buf, or tensor arenas.
1,890,168 AXI0 write beats: The number of AXI beats with write transactions to AXI0 bus.
1,740,069 AXI1 read beats: The number of AXI beats with read transactions from the AXI1 bus. AXI1 is the bus where the Ethos-U55 NPU reads the model. So, read-only.
For FPGA platforms, a CPU cycle count can also be enabled. However, do not use cycle counters for FVP, as the CPU model is not cycle-approximate or cycle-accurate.