To make a smart speaker

中文

Here is a collection of resources to make a smart speaker. Hope we can make an open source one for daily use. I believe we have enough resources to make an open source smart speaker. Let's do it. Take a look at the progress of the project named smart speaker from scratch on hackaday. The first hardware kit is available now.

The simplified flowchart of a smart speaker is like:

+---+   +----------------+   +---+   +---+   +---+
|Mic|-->|Audio Processing|-->|KWS|-->|STT|-->|NLU|
+---+   +----------------+   +---+   +---+   +-+-+
                                               |
                                               |
+-------+   +---+   +----------------------+   |
|Speaker|<--|TTS|<--|Knowledge/Skill/Action|<--+
+-------+   +---+   +----------------------+

• Audio Processing includes Acoustic Echo Cancellation (AEC), Beamforming, Noise Suppression (NS), etc.
• Keyword Spotting (KWS) detects a keyword (such as OK Google, Hey Siri) to start a conversation.
• Speech To Text (STT)
• Natural Language Understanding (NLU) converts raw text into structured data.
• Knowledge/Skill/Action - Knowledge base and plugins (Alexa Skill, Google Action) to provide an answer.
• Text To Speech

KWS + STT + NLU + Skill + TTS

Active open source projects

• Snips :star: - the first 100% on-device and private-by-design open-source Voice AI platform
• Mycroft :star: - a hackable open source voice assistant
• SEPIA :robot: - Highly customizable, open-source, cross-platform voice assistant and VUI framework (HTML + Java + x)
• Kalliope - a framework that will help you to create your own personal assistant, kind of similar with Mycroft (Both written by Python)
• dingdang robot - a :cn: voice interaction robot based on Jasper and built with raspberry pi

SDK

• Amazon Alexa Voice Service - is the most widely used voice assistant

  • C++ SDK
  • Java Client
  • Python Client

• Google Assistant SDK

  It has the smartest brain, its extension called Google Action can be created on a few steps with digitalflow.ai and its Device Action is very suit for home smart devices.

• Baidu DuerOS

• Snips

  • Install Snips on Raspberry Pi 3, Linux, osX, iOS and Android

• SEPIA Installation, SEPIA with Porcupine + ReSpeaker

KWS

• Mycroft Precise - A lightweight, simple-to-use, RNN wake word listener
• Snowboy - DNN based hotword and wake word detection toolkit
• Honk - PyTorch reimplementation of Google's TensorFlow CNNs for keyword spotting
• ML-KWS-For-MCU - Maybe the most promise for resource constrained devices such as ARM Cortex M7 microcontroller
• Porcupine - Lightweight, cross-platform engine to build custom wake words in seconds

STT

• Mozilla DeepSpeech - A TensorFlow implementation of Baidu's DeepSpeech architecture
• Kaldi
• wav2letter++ - a fast, open source speech processing toolkit from the Speech team at Facebook AI Research built to facilitate research in end-to-end models for speech recognition.
• Zamia Speech - Open tools, data, models (kaldi models and wav2letter++ models) for cloudless automatic speech recognition. It can be run on Raspberry Pi
• PocketSphinx - a lightweight speech recognition engine using HMM + GMM

NLU

• Rasa NLU

  • Rasa NLU for Chinese

• Snips NLU - a Python library that allows to parse sentences written in natural language and extracts structured information.

TTS

• Mozilla TTS - Deep learning for Text to Speech
• Mimic - Mycroft's TTS engine, based on CMU's Flite (Festival Lite)
• manytts - an open-source, multilingual text-to-speech synthesis system written in pure java
• espeak-ng - an open source speech synthesizer that supports 99 languages and accents.
• ekho - Chinese text-to-speech engine
• WaveNet, Tacotron 2

Audio Processing

• Acoustic Echo Cancellation

  • SpeexDSP, its python binding speexdsp-python
  • EC - Echo Cancelation Daemon based on SpeexDSP AEC for Raspberry Pi or other devices running Linux.

• Direction Of Arrival (DOA) - Most used DOA algorithms is GCC-PHAT

  • tdoa
  • odas - ODAS stands for Open embeddeD Audition System. This is a library dedicated to perform sound source localization, tracking, separation and post-filtering. ODAS is coded entirely in C, for more portability, and is optimized to run easily on low-cost embedded hardware. ODAS is free and open source.

• Beamforming

  • BeamformIt - filter&sum beamforming
  • CGMM Beamforming - a reference implementation
  • MVDR Beamforming
  • GSC Beamforming

• Voice Activity Detection

  • WebRTC VAD, py-webrtcvad
  • DNN VAD

• Noise Suppresion

  • NS of WebRTC audio processing, python-webrtc-audio-processing

Audio I/O

• PortAudio, pyaudio
• libsoundio
• ALSA
• PulseAudio
• Pipewire