Bonus: Quick Dip Into The Technology

So how do these products work? 

The first component of automatic speech recognition is, of course, speech. Speech must be converted from physical sound to an electrical signal with a microphone, and then to digital data with an analog-to-digital converter. It sounds complicated but this is only the easy part!

IBM at INTERSPEECH 2019 sum up the next steps in Voice AI succinctly: traditional "hybrid" ASR systems are comprised of an acoustic model, language model, and pronunciation model. Each component requires complex training. This is different from the current research into Ensemble systems that we will leave aside for now. 

What are these models? The acoustic model is a multi-stage process of model training and time alignment. Using a Hidden Markov Model (HMM) the model chops the speech into short snippets of about ten milliseconds. Because we assume each snippet is "stationary" (nothing really changes over the 10 milli-seconds), we can essentially create and plot the "power" of each snippet as some function of frequency. In other words, we can describe the qualities or attributes of that snippet in a group or vector of numbers. The number of attributes in each vector is usually small - sometimes as low as 10, although more accurate systems may have 32 or more. The final output of the HMM is a sequence of these vectors, just a long table of numbers describing each snippet and models that match these snippets to one or more phonemes, the fundamental unit of speech. This calculation requires training, since the sound of a phoneme varies from speaker to speaker, and even varies from one utterance to another by the same speaker. The model uses a special algorithm that can guess the most likely word(s) that produce the given sequence of phonemes. These groups of words are then processed by a Language Model that guess the next word(s) in a sequence, tries to guess intent, sentiment or other Natural Language Processing tasks. Delvify is adept at using the newest BERT models.

As you can imagine this whole process requires much specialized work. In most modern speech recognition systems, there are many shortcuts using neural networks to simplify the speech signal before the HMM recognition. Voice activity detectors (VADs) are also used to reduce an audio signal to only the portions that are likely to contain speech. This prevents the recognizer from wasting time analyzing unnecessary parts of the signal such as background noise. This is not the same as noise cancellation headphones!