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 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 and if a set keyword is detected then an Automatic Speech Recognition model is run on Ethos-U55 on the remaining audio. Tensor arena memory region is reused between models to optimise application memory footprint.
"Yes" key word is used to trigger full command recognition following the key word. Use case code could be found in source/use_case/kws_asr directory.
In this use-case there are 2 different models being used with different requirements for preprocessing. As such each preprocessing process is detailed below. Note that Automatic Speech Recognition only occurs if a keyword is detected in the audio clip.
By default the KWS model is run purely on CPU and not on the Ethos-U55.
The DS-CNN keyword spotting 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 extracted from audio data and can be used as input for machine learning tasks like keyword spotting and speech recognition. See source/application/main/include/Mfcc.hpp for implementation details.
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 needed to perform an inference is reached. In total this will be 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 (i.e. do an inference for samples 8000-24000).
Note: Parameters of the MFCC feature extraction such as window size, stride, number of features etc. all depend on what was used during model training. These values are specific to each model and if you try a different keyword spotting model that uses MFCC input then values are likely to need changing to match the new model.
In addition, MFCC feature extraction methods can vary slightly with different normalization methods or scaling etc. 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 extracted from audio data and can be used as input for machine learning tasks like keyword spotting and speech recognition. See source/application/main/include/Mfcc.hpp for implementation details.
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 will be 296 windows that each have 13 MFCC features calculated for them.
After extracting MFCC features the first and second order derivatives of these features with respect to 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 will have a shape of 296x39.
These extracted features are quantized, and an inference is performed.
For longer audio clips where multiple inferences need to 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 1 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 the MFCC features will be padded by repeating the last calculated feature until an inference can be performed.
Note: Parameters of the MFCC feature extraction such as window size, stride, number of features etc. 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 audio samples we offset by 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 needs to 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 only certain parts of the output are usable depending on the bit of the audio clip that is currently being processed.
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 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 output before continuing.
Once any rows have been removed the final processing can be done. To process the output, first the letter with the highest probability at each time step is found. Next, any letters that are repeated multiple times in a row are removed (e.g. [t, t, t, o, p, p] becomes [t, o, p]). Finally, the 29^th^ blank token letter is removed from the output.
For the final output, the result 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, Keyword Spotting and Automatic Speech Recognition use case adds:
kws_asr_MODEL_TFLITE_PATH_ASR and kws_asr_MODEL_TFLITE_PATH_KWS: Path to the NN model files in TFLite format. Models will be processed and 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_U55_ENABLED should be aligned with the chosen model, i.e:
ETHOS_U55_ENABLED is set to On or 1, the NN model is assumed to be optimized. The model will naturally fall back to the Arm® Cortex®-M CPU if an unoptimized model is supplied.ETHOS_U55_ENABLED is set to Off or 0, the NN model is assumed to be unoptimized. Supplying an optimized model in this case will result in a runtime error.kws_asr_FILE_PATH: 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 containing the delivered set of audio clips.
kws_asr_LABELS_TXT_FILE_KWS and kws_asr_LABELS_TXT_FILE_ASR: Path respectively to 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: Input data sampling rate. Each audio file from kws_asr_FILE_PATH is preprocessed during the build to match NN model input requirements. Default value is 16000.
kws_asr_AUDIO_MONO: If set to ON the audio data will be converted to mono. Default is ON.
kws_asr_AUDIO_OFFSET: Start loading audio data starting from this offset (in seconds). Default value is 0.
kws_asr_AUDIO_DURATION: Length of the audio data to be used in the application in seconds. Default is 0 meaning the whole audio file will be taken.
kws_asr_AUDIO_MIN_SAMPLES: Minimum number of samples required by the network model. If the audio clip is shorter than this number, it is padded with zeros. Default value is 16000.
kws_asr_MODEL_SCORE_THRESHOLD_KWS: Threshold value that must be applied to the keyword spotting inference results for a label to be deemed valid. 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. Default is 0.5.
kws_asr_ACTIVATION_BUF_SZ: The intermediate/activation buffer size reserved for the NN model. By default, it is set to 2MiB and should be enough for most models.
In order to build ONLY Keyword Spotting and Automatic Speech Recognition example application add to the cmake command line specified in Building -DUSE_CASE_BUILD=kws_asr.
Note: This section describes the process for configuring the build for
MPS3: SSE-300for different target platform see Building.
Create a build directory and navigate inside:
mkdir build_kws_asr && cd build_kws_asr
On Linux, execute the following command to build the application to run on the Ethos-U55 Fast Model when providing only the mandatory arguments for CMake configuration:
cmake \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=./scripts/cmake/bare-metal-toolchain.cmake \ -DUSE_CASE_BUILD=kws_asr ..
Toolchain option CMAKE_TOOLCHAIN_FILE points to the toolchain specific file to set the compiler and platform specific parameters.
To configure a build that can be debugged using Arm-DS, we can just specify the build type as Debug:
cmake \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DCMAKE_BUILD_TYPE=Debug \ -DUSE_CASE_BUILD=kws_asr ..
To configure a build that can be debugged using a tool that only supports DWARF format 3 (Modeldebugger for example), we can use:
cmake \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DCMAKE_BUILD_TYPE=Debug \ -DARMCLANG_DEBUG_DWARF_LEVEL=3 \ -DUSE_CASE_BUILD=kws_asr ..
Note: If building for different Ethos-U55 configurations, see Configuring build for different Arm Ethos-U55 configurations:
If the TensorFlow source tree is not in its default expected location, set the path using TENSORFLOW_SRC_PATH. Similarly, if the Ethos-U55 driver is not in the default location, ETHOS_U55_DRIVER_SRC_PATH can be used to configure the location. For example:
cmake \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DTENSORFLOW_SRC_PATH=/my/custom/location/tensorflow \ -DETHOS_U55_DRIVER_SRC_PATH=/my/custom/location/core_driver \ -DUSE_CASE_BUILD=kws_asr ..
Also, CMSIS_SRC_PATH parameter can be used to override the CMSIS sources used for compilation used by TensorFlow by default. For example, to use the CMSIS sources fetched by the ethos-u helper script, we can use:
cmake \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DTENSORFLOW_SRC_PATH=../ethos-u/core_software/tensorflow \ -DETHOS_U55_DRIVER_SRC_PATH=../ethos-u/core_software/core_driver \ -DCMSIS_SRC_PATH=../ethos-u/core_software/cmsis \ -DUSE_CASE_BUILD=kws_asr ..
Note: If re-building with changed parameters values, it is highly advised to clean the build directory and re-run the CMake command.
If the CMake command succeeded, build the application as follows:
make -j4
For Windows, use mingw32-make.
Add VERBOSE=1 to see compilation and link details.
Results of the build will be placed under build/bin folder:
bin ├── ethos-u-kws_asr.axf ├── ethos-u-kws_asr.htm ├── ethos-u-kws_asr.map ├── images-kws_asr.txt └── sectors └── kws_asr ├── dram.bin └── itcm.bin
Where:
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 (e.g. libraries used, what was optimized, location of objects)
ethos-u-kws_asr.htm: Human readable file containing the call graph of application functions.
sectors/: Folder containing the built application, split into files for loading into different FPGA memory regions.
Images-kws_asr.txt: Tells the FPGA which memory regions to use for loading the binaries in sectors/** folder.
The application performs inference on data found in the folder set by the CMake 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 files into this folder:
mkdir /tmp/custom_files cp custom_audio1.wav /tmp/custom_files/
Note: Clean the build directory before re-running the CMake command.
Next set kws_asr_FILE_PATH to the location of this folder when building:
cmake \ -Dkws_asr_FILE_PATH=/tmp/custom_files/ \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DUSE_CASE_BUILD=kws_asr- ..
The files found in the kws_asr_FILE_PATH folder will be picked up and automatically converted to C++ files during the CMake configuration stage and then compiled into the application during the build phase for performing inference with.
The log from the configuration stage should tell you what directory path has been used:
-- User option kws_asr_FILE_PATH is set to /tmp/custom_files
After compiling, your custom inputs will 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 and ASR inference using the model pointed to by the CMake parameter kws_asr_MODEL_TFLITE_PATH_ASR.
This section assumes you wish to change the existing ASR model to a custom one. If instead you wish to change the KWS model then the instructions will be the same except ASR will change to KWS.
Note: If you want to run the model using Ethos-U55, ensure your custom model has been run through the Vela compiler successfully before continuing. See Optimize model with Vela compiler.
To run the application with a custom model you will need to provide a labels_<model_name>.txt file of labels associated with the model. Each line of the file should correspond to one of the outputs in your model. See the provided labels_wav2letter.txt file for an example.
Then, you must set kws_asr_MODEL_TFLITE_PATH_ASR to the location of the Vela processed model file and kws_asr_LABELS_TXT_FILE_ASR to the location of the associated labels file.
An 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> \ -DTARGET_PLATFORM=mps3 \ -DTARGET_SUBSYSTEM=sse-300 \ -DCMAKE_TOOLCHAIN_FILE=scripts/cmake/bare-metal-toolchain.cmake \ -DUSE_CASE_BUILD=kws_asr ..
For Windows, add -G "MinGW Makefiles" to the CMake command.
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, labels text files pointed to by kws_asr_LABELS_TXT_FILE_KWS and kws_asr_LABELS_TXT_FILE_ASR will be converted to C++ files during the CMake configuration stage and then compiled into the application for performing inference with.
The log from the configuration stage should tell you what model path and labels file have been used:
-- 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 will have now replaced the default one in the application.
The FVP is available publicly from Arm Ecosystem FVP downloads.
For Ethos-U55 evaluation, please download the MPS3 version of the Arm® Corstone™-300 model that contains Ethos-U55 and Cortex-M55. The model is currently only supported on Linux based machines. To install the FVP:
Unpack the archive
Run the install script in the extracted package
./FVP_Corstone_SSE-300_Ethos-U55.sh
Once completed the building step, application binary ethos-u-kws_asr.axf can be found in the build/bin folder. Assuming the install location of the FVP was set to ~/FVP_install_location, the simulation can be started by:
$ ~/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 should appear 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 will also launch a telnet window with the sample application's standard output and error log entries containing information about the pre-built application version, TensorFlow Lite Micro library version used, data type as well as the input and output tensor sizes of the model compiled into the executable binary.
After the application has started if kws_asr_FILE_PATH pointed to a single file (or a folder containing a single input file) the inference starts immediately. In case of multiple inputs choice, it outputs a menu and waits for the user input from telnet terminal:
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:
“Classify next audio clip” menu option will run single inference on the next included file.
“Classify audio clip at chosen index” menu option will run inference on the chosen audio clip.
Note: Please make sure to select audio clip index in the range of supplied audio clips during application build.
“Run ... on all” menu option triggers sequential inference executions on all built-in files.
“Show NN model info” menu option prints information about model data type, input and output tensor sizes:
INFO - uTFL version: 2.5.0 INFO - Model INPUT tensors: INFO - tensor type is INT8 INFO - tensor occupies 490 bytes with dimensions INFO - 0: 1 INFO - 1: 1 INFO - 2: 49 INFO - 3: 10 INFO - Quant dimension: 0 INFO - Scale[0] = 1.107164 INFO - ZeroPoint[0] = 95 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] = 0.003906 INFO - ZeroPoint[0] = -128 INFO - Activation buffer (a.k.a tensor arena) size used: 123616 INFO - Number of operators: 16 INFO - Operator 0: RESHAPE INFO - Operator 1: CONV_2D INFO - Operator 2: DEPTHWISE_CONV_2D INFO - Operator 3: CONV_2D INFO - Operator 4: DEPTHWISE_CONV_2D INFO - Operator 5: CONV_2D INFO - Operator 6: DEPTHWISE_CONV_2D INFO - Operator 7: CONV_2D INFO - Operator 8: DEPTHWISE_CONV_2D INFO - Operator 9: CONV_2D INFO - Operator 10: DEPTHWISE_CONV_2D INFO - Operator 11: CONV_2D INFO - Operator 12: AVERAGE_POOL_2D INFO - Operator 13: RESHAPE INFO - Operator 14: FULLY_CONNECTED INFO - Operator 15: SOFTMAX 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: 809808 INFO - Number of operators: 1 INFO - Operator 0: ethos-u
“List” menu option prints a list of pair ... indexes - the original filenames embedded in the application:
[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 application output:
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: 0.996094 INFO - Profile for Inference: INFO - NPU AXI0_RD_DATA_BEAT_RECEIVED beats: 217385 INFO - NPU AXI0_WR_DATA_BEAT_WRITTEN beats: 82607 INFO - NPU AXI1_RD_DATA_BEAT_RECEIVED beats: 59608 INFO - NPU ACTIVE cycles: 680611 INFO - NPU IDLE cycles: 561 INFO - NPU total cycles: 681172 INFO - Keyword spotted INFO - Inference 1/2 INFO - Inference 2/2 INFO - Result for inf 0: no gow INFO - Result for inf 1: stoppe INFO - Final result: no gow stoppe INFO - Profile for Inference: INFO - NPU AXI0_RD_DATA_BEAT_RECEIVED beats: 13520864 INFO - NPU AXI0_WR_DATA_BEAT_WRITTEN beats: 2841970 INFO - NPU AXI1_RD_DATA_BEAT_RECEIVED beats: 2717670 INFO - NPU ACTIVE cycles: 28909309 INFO - NPU IDLE cycles: 863 INFO - NPU total cycles: 28910172
It could take several minutes to complete one inference run (average time is 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-U55's PMU report:
28,910,172 total cycle: The number of NPU cycles
28,909,309 active cycles: number of NPU cycles that were used for computation
863 idle cycles: number of cycles for which the NPU was idle
13,520,864 AXI0 read beats: The number of AXI beats with read transactions from AXI0 bus. AXI0 is the bus where Ethos-U55 NPU reads and writes to the computation buffers (activation buf/tensor arenas).
2,841,970 AXI0 write beats: The number of AXI beats with write transactions to AXI0 bus.
2,717,670 AXI1 read beats: The number of AXI beats with read transactions from AXI1 bus. AXI1 is the bus where Ethos-U55 NPU reads the model (read only)
For FPGA platforms, CPU cycle count can also be enabled. For FVP, however, CPU cycle counters should not be used as the CPU model is not cycle-approximate or cycle-accurate.
Note that in this example the KWS inference does not use the Ethos-U55 and is run purely on CPU, therefore 0 Active NPU cycles is shown.