[SpeechCommands] Clean up training/browser-fft directory (#513)

- Removes obsolete Python files, now that we have notebooks showing the complete workflow
- Removes supporting TypeScript scripts and package.json file.
- Updates README.md.
- Also adds the step to write meta

Author: caisq

run_python_tests.sh

@@ -20,4 +20,5 @@ set -e
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
-"${SCRIPT_DIR}/speech-commands/training/browser-fft/py_lint_and_test.sh"
+# This file is currently an empty placeholder. Add Python tests here when
+# they are added in the future.

speech-commands/training/browser-fft/README.md

@@ -1,85 +1,22 @@
 # Training a TensorFlow.js model for Speech Commands Using Browser FFT
 
-## Preparing data for training
-
-Before you can train your model that uses spectrogram from the browser's
-WebAudio as input features, you need to convert the speech-commands
-data set into a format that TensorFlow.js can ingest, by running the
-data through the native WebAudio frequency analyzer (FFT) of the browser.
-The following steps are involved:
-
-1. Download the speech-commands data set from
-   https://storage.cloud.google.com/download.tensorflow.org/data/speech_commands_v0.01.tar.gz
-   or
-   https://storage.cloud.google.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz
-   Version 0.02 is a larger dataset than 0.01.
-
-2. Use `prep_wavs.py` to convert the raw wav files into a binary format
-   ready for FFT conversion in the browser. E.g.,
-
-   ```sh
-   python prep_wavs.py \
-      --words zero,one,two,three,four,five,six,seven,eight,nine,go,stop,left,right,up,down \
-      --test_split 0.15 \
-      --include_noise \
-       "${HOME}/ml-data/speech_commands_data" \
-       "${HOME}/ml-data/speech_commands_data_converted"
-   ```
-
-   With the `--words` flag, you can specify what words to include in the
-   training of the model. With the `--test_split` flag, you can specify the
-   fraction of the .wav files that will be randomly drawn for testing after
-   training. The `--include_noise` flag asks the script to randomly draw
-   segments from the long .wav files in the '_background_noise_' folder to
-   generate training (and test) examples for background noise. (N.B.: this is
-   *not* about adding noise to the word examples.)
-   The last two arguments point to the input and output directories,
-   respectively.
-
-   Under the output path (i.e., `speech_commands_data_converted` in this example),
-   there will be two subfolders, called `train` and `test`, which hold the
-   training and testing splits, respectively. Under each of `train` and `test`,
-   there are subfolders with names matching the words (e.g., `zero`, `one`,
-   etc.) In each of those subfolders, there will subfolders with names
-   such as `0` and `1`, which contain a number of `.dat` files.
-
-3. Run WebAudio FFT on the `.dat` files generated in step 2 in the browser.
-   TODO(cais): Provide more details here.
-
-## Training the TensorFlow.js Model in tfjs-node or tfjs-node-gpu
-
-1. Download and extract the browser-FFT version of the speech-commands dataset:
-
-   ```sh
-   curl -fSsL https://storage.googleapis.com/learnjs-data/speech-commands/speech-commands-data-v0.02-browser.tar.gz  -o speech-commands-data-v0.02-browser.tar.gz && \
-   tar xzvf speech-commands-data-v0.02-browser.tar.gz
-   ```
-
-2. Start training. First, download JavaScript dependencies using:
-
-   ```sh
-   yarn
-   ```
-
-   Then, to train the model using CPU (tfjs-node):
-
-   ```sh
-   yarn train speech-commands-data-v0.02-browser/ ./my-model/
-   ```
-
-   Or, to train the model using a GPU (tfjs-node-gpu,
-   requires CUDA-enabled GPU and drivers):
-
-   ```sh
-   yarn train --gpu speech-commands-data-v0.02-browser/ ./my-model/
-   ```
-
-## Development
-
-### Python
-
-To run linting and tests of the Python files in this directory, use script:
-
-```sh
-./py_lint_and_test.sh
-```
+This directory contains two example notebooks. They demonstrate how to train
+custom TensorFlow.js audio models and deploy them for inference. The models
+trained this way expect inputs to be spectrograms in a format consistent with
+[WebAudio's `getFloatFrequencyData`](https://developer.mozilla.org/en-US/docs/Web/API/AnalyserNode/getFloatFrequencyData).
+Therefore they can be deployed to the browser using the speech-commands library
+for inference.
+
+Specifically,
+
+- [training_custom_audio_model_in_python.ipynb](./training_custom_audio_model_in_python.ipynb)
+  contains steps to preprocess a directory with audio examples stored as .wav
+  files and the steps in which a tf.keras model can be trained on the
+  preprocessed data. It then demonstrates how the trained tf.keras model can be
+  converted to a TensorFlow.js `LayersModel` that can be loaded with the
+  speech-command library's `create()` API. In addition, the notebook also shows
+  the steps to convert the trained tf.keras model to a TFLite model for
+  inference on mobile devices.
+- [tflite_conversion.ipynb](./tflite_conversion.ipynb) illustrates how
+  an audio model trained on [Teachable Machine](https://teachablemachine.withgoogle.com/train/audio)
+  can be converted to TFLite directly.