Build your own voice ai. This repo is for my YouTube video series on building an AI voice assistant with PyTorch.

Looking for contributors to help build out the assistant. There is still alot of work to do. This would be a good oppurtunity to learn Machine Learning and how to Engineer an entire ML system from the ground up. If you're interested join the Discord Server

TODO:

• wake word model and engine
• pre-trained wake word model use for fine tuning on your own wakeword
• speech recognition model, pretrained model, and engine
• natural langauge understanding model, pretrained model, and engine
• speech synthesis model, pretrained model, and engine
• skills framework
• Core A.I. Voice Assistant logic to integrate wake word, speech recongition, natural language understanding, speech sysnthesis, and the skills framework.

• python3
• portaudio (for recording with pyaudio to work)
• ctcdecode - for speechrecognition

If you're on mac you can install portaudio using homebrew

NOTICE: If you are using windows, some things may not work. For example, torchaudio. I suggest trying this on linux or mac, or use wsl2 on windows

1. virtualenv voiceassistant.venv
2. source voiceassistant.venv/bin/activate

pip install -r requirements.txt

If you are running with just the cpu docker build -f cpu.Dockerfile -t voiceassistant .

If you are running on a cuda enabled machine docker build -f Dockerfile -t voiceassistant .

Youtube Video For WakeWord

For more details make sure to visit these files to look at script arguments and description

wakeword/neuralnet/train.py is used to train the model

wakeword/neuralnet/optimize_graph.py is used to create a production ready graph that can be used in engine.py

wakeword/engine.py is used to demo the wakeword model

wakeword/scripts/collect_wakeword_audio.py - used to collect wakeword and environment data

wakeword/scripts/split_audio_into_chunks.py - used to split audio into n second chunks

wakeword/scripts/split_commonvoice.py - if you download the common voice dataset, use this script to split it into n second chunks

wakeword/scripts/create_wakeword_jsons.py - used to create the wakeword json for training

For more details make sure to visit these files to look at script arguments and description

1. collect data

   1. environment and wakeword data can be collected using python collect_wakeword_audio.py

      cd VoiceAssistant/wakeword/scripts
      mkdir data
      cd data
      mkdir 0 1 wakewords
      python collect_wakeword_audio.py --sample_rate 8000 --seconds 2 --interactive --interactive_save_path ./data/wakewords
      

   2. to avoid the imbalanced dataset problem, we can duplicate the wakeword clips with

      python replicate_audios.py --wakewords_dir data/wakewords/ --copy_destination data/1/ --copy_number 100
      

   3. be sure to collect other speech data like common voice. split the data into n seconds chunk with split_audio_into_chunks.py.
   4. put data into two seperate directory named 0 and 1. 0 for non wakeword, 1 for wakeword. use create_wakeword_jsons.py to create train and test json
   5. create a train and test json in this format...

      // make each sample is on a seperate line
      {"key": "/path/to/audio/sample.wav, "label": 0}
      {"key": "/path/to/audio/sample.wav, "label": 1}
      

2. train model

   1. use train.py to train model
   2. after model training us optimize_graph.py to create an optimized pytorch model

3. test

   1. test using the engine.py script

YouTube Video for Speech Recognition

For more details make sure to visit these files to look at script arguments and description

speechrecognition/scripts/mimic_create_jsons.pyis used to create the train.json and test.json files with Mimic Recording Studio

speechrecognition/scripts/commonvoice_create_jsons.pyis used to convert mp3 into wav and create the train.json and test.json files with the Commonvoice dataset

spechrecognition/neuralnet/train.py is used to train the model

spechrecognition/neuralnet/optimize_graph.py is used to create a production ready graph that can be used in engine.py

spechrecognition/engine.py is used to demo the speech recognizer model

spechrecognition/demo/demo.py is used to demo the speech recognizer model with a Web GUI

The pretrained model can be found here at this google drive

1. Collect your own data - the pretrain model was trained on common voice. To make this model work for you, you can collect about an hour or so of your own voice using the Mimic Recording Studio. They have prompts that you can read from.

   1. collect data using mimic recording studio, or your own dataset.
   2. be sure to chop up your audio into 5 - 16 seconds chunks max.
   3. create a train and test json in this format...

       // make each sample is on a seperate line
       {"key": "/path/to/audio/speech.wav, "text": "this is your text"}
       {"key": "/path/to/audio/speech.wav, "text": "another text example"}
   

   use mimic_create_jsons.py to create train and test json's with the data from Mimic Recording Studio.

    python mimic_create_jsons.py --file_folder_directory /dir/to/the/folder/with/the/studio/data --save_json_path /path/where/you/want/them/saved
   

   (The Mimic Recording Studio files are usually stored in ~/mimic-recording-studio-master/backend/audio_files/[random_string].)

   use commonvoice_create_jsons.py to convert from mp3 to wav and to create train and test json's with the data from Commonvoice by Mozilla

    python commonvoice_create_jsons.py --file_path /path/to/commonvoice/file/.tsv --save_json_path /path/where/you/want/them/saved 
   

   if you dont want to convert use --not-convert

2. Train model

   1. use train.py to fine tune. checkout the train.py argparse for other arguments

      python train.py --train_file /path/to/train/json --valid_file /path/to/valid/json --load_model_from /path/to/pretrain/speechrecognition.ckpt
   

   2. To train from scratch omit the --load_model_from argument in train.py
   3. after model training us optimize_graph.py to create a frozen optimized pytorch model. The pretrained optimized torch model can be found in the google drive link as speechrecognition.zip

3. test

   1. test using the engine.py script

documenation to get this running on rpi is in progress...