Added code

.github/workflows/build.yml

@@ -0,0 +1,36 @@
+name: build
+
+on:
+  push:
+    branches:    
+      - build
+jobs:
+  build:
+
+    env:
+      BUILD_TYPE: Release
+
+    runs-on: ubuntu-latest
+
+    steps:
+
+    - name: Prepare
+      run: |
+        sudo apt-get update
+
+    - name: Checkout
+      uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+
+    - name: Build
+      run: |
+        mkdir build
+        cd build
+        cmake -DCMAKE_BUILD_TYPE=$BUILD_TYPE ..
+        make
+
+    - uses: actions/upload-artifact@0b2256b8c012f0828dc542b3febcab082c67f72b
+      with:
+        name: binaries
+        path: |
+          build/kws_cli
+

.gitignore

@@ -1,11 +1 @@
-CMakeLists.txt.user
-CMakeCache.txt
-CMakeFiles
-CMakeScripts
-Testing
-Makefile
-cmake_install.cmake
-install_manifest.txt
-compile_commands.json
-CTestTestfile.cmake
-_deps
+build/

CMakeLists.txt

@@ -0,0 +1,16 @@
+cmake_minimum_required(VERSION 3.16.3)
+
+project(kws_cli)
+
+set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -s")
+
+include_directories(include)
+set(LIBS m)
+
+add_compile_options(-Wno-deprecated-declarations
+                    -Wall -Werror -fPIC)
+
+add_executable(kws_cli src/kws_cli.c src/fbank.c)
+target_link_libraries(kws_cli ${LIBS})
+
+

README.md

@@ -1 +1,51 @@
-# kws_cli
+# kws_cli
+
+[![build](https://github.com/mryndzionek/kws_cli/actions/workflows/build.yml/badge.svg)](https://github.com/mryndzionek/kws_cli/actions/workflows/build.yml)
+
+## About
+
+Speech recognition in ~300kB of code.
+
+Small footprint, standalone, zero dependency, offline
+keyword spotting (KWS) CLI tool. Might be useful in
+some automation task. Accepts audio on stdin a and recognizes
+following words: `up`, `down`, `left`, `right`, `stop`.
+
+Here is an example invocation:
+
+```
+rec -q -t raw -c1 -e signed -b 16 -r16k - | ./kws_cli
+```
+
+Make sure you have decent microphone and the system audio
+is on a decent level.
+
+Individual WAV files can piped (e.g. for testing) using:
+
+```
+sox -S ../untitled.wav -t raw -c1 -e signed -b 16 -r16k - | ./kws_cli
+```
+
+## More details
+
+Speech recognition is based on [this](https://github.com/microsoft/EdgeML/blob/master/docs/publications/Sha-RNN.pdf)
+model and examples from the same repository.
+This simple model with three layers: 2x LSTM + 1x fully connected.
+The model is trained in PyTorch and exported to ONNX.
+Then [onnx2c](https://github.com/kraiskil/onnx2c)
+is used to convert the model to a bunch of C code.
+The LSTM layers had become mainstream in recent years and are well
+supported in different frameworks. The model is small, so it might
+be possible to run it on Cortex-M4/M7, or ESP32 (future work).
+
+## Building
+
+The usual CMake routine:
+
+```
+mkdir build
+cd build
+cmake -DCMAKE_BUILD_TYPE=Release
+make
+```
+

include/fbank.h

@@ -0,0 +1,14 @@
+#ifndef __FBANK_H__
+#define __FBANK_H__
+
+#define SAMPLE_LEN (16000UL)
+#define NUM_FRAMES (99UL)
+#define NUM_FILT (32UL)
+
+void fbank(float input[SAMPLE_LEN], float output[NUM_FRAMES][NUM_FILT]);
+void fbank_norm(float inputoutput[NUM_FILT]);
+void fbank_speech_detect(float input[NUM_FRAMES][NUM_FILT], size_t *label, float *logit);
+void fbank_print_min_max(float input[NUM_FRAMES][NUM_FILT]);
+char const *const fbank_label_idx_to_str(size_t label);
+
+#endif // __FBANK_H__

src/kws_cli.c

@@ -0,0 +1,77 @@
+#include <stdio.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#include <signal.h>
+#include <string.h>
+
+#include "fbank.h"
+
+#define CHUNK_SIZE (SAMPLE_LEN / 2)
+
+static volatile bool byebye = false;
+
+static void INThandler(int sig)
+{
+    byebye = true;
+}
+
+int main(int argc, char *argv[])
+{
+    int16_t in[CHUNK_SIZE] = {0.0};
+    float input[SAMPLE_LEN] = {0.0};
+    float fbins[NUM_FRAMES][NUM_FILT];
+    size_t n = 0;
+    size_t label;
+    float logit;
+    bool debounce_active = false;
+
+    struct sigaction sa = {.sa_handler = INThandler, .sa_flags = 0};
+    sigaction(SIGINT, &sa, 0);
+
+    fprintf(stderr, "\nStarting...\n");
+
+    while (!byebye)
+    {
+        size_t m = fread(&in[n], sizeof(int16_t), CHUNK_SIZE - n, stdin);
+        n += m;
+
+        if (m == 0)
+        {
+            break;
+        }
+
+        if (n == CHUNK_SIZE)
+        {
+            if (CHUNK_SIZE < SAMPLE_LEN)
+            {
+                memmove(input, &input[CHUNK_SIZE], (SAMPLE_LEN - CHUNK_SIZE) * sizeof(float));
+            }
+
+            for (size_t i = 0; i < CHUNK_SIZE; i++)
+            {
+                input[SAMPLE_LEN - CHUNK_SIZE + i] = in[i];
+            }
+
+            fbank(input, fbins);
+            fbank_speech_detect(fbins, &label, &logit);
+            if (debounce_active)
+            {
+                if (label == 0)
+                {
+                    debounce_active = false;
+                }
+            }
+            else
+            {
+                if (label > 0)
+                {
+                    fprintf(stderr, "label: '%s', label_idx: %ld, logit: %f\n", fbank_label_idx_to_str(label), label, logit);
+                    fflush(stderr);
+                    debounce_active = true;
+                }
+            }
+            n = 0;
+        }
+    }
+}