Figure 2: Development of a voice dialog.

© Hy-Line

Speech synthesis turns voice control with a focus on voice input and voice output, even for extensive texts, into a fully comprehensive assistance system. This allows the operator or service technician to select relevant text passages from a stored operating manual using suitable search terms and have them read aloud. During troubleshooting, the eyes remain focused on the machine.

AI is used to create the synthesis models for text-to-speech (TTS): the models use machine learning algorithms that help to convert continuous text into dynamic, natural-sounding speech output. As with speech recognition training, the process here is also two-stage: training in the cloud, interpretation and playback only locally - so data remains confidential and secure.

Kickstart for professional voice control

The main medium is still manual input - whether using a keyboard, mouse, gesture control or control buttons. Voice can replace input wherever hands are not available because they are being used for other purposes or are dirty. This includes, for example, the HMI on the machine in the production line, where both hands are needed for the workpiece, or the information system at the point of sale, which provides information on where to find stores in the shopping mall or products on the shelves. In the catering industry, the temperature of professional kitchen appliances can be set to the exact degree, while the hands remain clean for the food. In logistics, the storage system gives instructions on where an item should be picked or placed. In medical technology, it is important to keep hands sterile or not to contaminate them so that viruses and bacteria are not passed on. New fields such as smart caravanning are also suitable for voice control: Where individual solutions are currently used for switching lights or querying the filling of fresh or service water tanks, a uniform interface with voice control can ensure simpler wiring and more ergonomic operation.

A ready-made hardware and software solution paves the way from the idea to the finished implementation of voice control. Figure 1 shows the starter kit, which not only makes the first steps easier. In order to develop a device that meets professional requirements and can be used around the clock, a WebSDK is available that abstracts the required algorithms and models.

Different languages are already stored in modules. The developer creates the SUI for the individual application with specific dialogs and commands. Below this is the machine interface, which passes on commands from the SUI to the hardware and GUI. To make this process as simple as possible, Hy-Line has developed a starter kit that not only simplifies the first steps on the way to a commercial solution.

The software

A web SDK is available as part of the starter kit, which can be used to explore the examples and create your own applications. You can create your own dialog models without programming by entering operating phrases with keywords and compiling them on the server. The result is downloaded to the starter kit and works without an Internet connection. The language system grows iteratively by formulating synonyms as alternative inputs and additional command phrases. The architecture receives the text and automatically recognizes keywords, which it assigns as subject or predicate. Filler words such as "please" and "uh" are skipped. The SDK provides APIs that can be transferred to the device via MQTT. This converts the recognized voice command into a hardware action. This reaction can be a voice output, a switching of a port, an output on the display or the change of a value in a JSON file. The kit is versatile enough to control external devices so that it can be used to create functional prototypes and test acceptance in the target group.

The hardware

Figure 3. runtime sequence.

© Hy-Line

The voice control kit is powered by a single-board computer in picoITX format based on the iMX8.M CPU. The operating interface is a 10.1-inch display with HD resolution and a capacitive touchscreen. All components are suitable for industrial use, so that a commercial implementation can be carried out with the starter kit. The application created in this way can also be ported to another target platform. In the simplest case, the acoustic output can be provided by a buzzer. However, it is better to use a loudspeaker that can emit broadband acknowledgement tones and voice messages. While earlier systems used to combine previously recorded audio snippets to output messages - such as announcing the time and date - TTS now offers the freedom to output any text in any language from a text file. The vocabulary is therefore practically unlimited and, just like voice input, works locally on the system without an internet connection at runtime.

Implementation process

Using a web-based development environment, the following steps are required to define a system for your own application. The voice dialog, i.e. the activation word with which the system's attention is drawn to the input, the permitted commands and their parameters, are compiled in the web tool as text input (Fig. 2). The first processing step takes place during input: Graphemes, i.e. characters entered, are converted into phonemes, i.e. the smallest acoustic components of speech.

Once all the words have been defined, the AI-based algorithms translate the defined language resources into a statistical and semantic model and offer them for download. The result is downloaded to the target platform and started. The network plug can then be pulled - the end product runs autonomously. The process in the finished application is shown in Figure 3.

The audio technology

The brain's ability to use two ears and the geometry of the head to isolate sounds and completely block out others is astonishing. This enables us to focus on the conversation with the other person at a table in a restaurant with lots of diners, while blocking out the neighbors who are also talking and the clattering of dishes. This is not so easy for a speech system: only with the help of a directional microphone or electronic filters can the system achieve an equally high recognition quality by increasing the signal-to-noise ratio. The directional microphone does not have to have the long design familiar from TV interviews. An array of several individual microphones makes it possible to identify the speaker of the "wake word" even in a noisy environment and to follow them if necessary. This increases the recognition accuracy, reaction speed and acceptance of the system enormously. The same technology can be used on the audio output side to emit sound in a specific direction.

Opening up new dimensions

The author: Rudolf Sosnowsky is Head of Technology at Hy-Line Computer Components Vertriebs GmbH in Unterhaching.

© Hy-Line

With the addition of language, every user interface gains a new dimension. Implementation is easier than expected, as the starter kit not only allows you to start a demo, but also to take your first steps with your own commands and outputs. A powerful SDK is available for implementing protocols to control external devices. State-of-the-art technology means that the system works independently of the speaker; 30 languages are predefined. This solution can also be used on platforms with limited CPU and memory resources; a digital signal processor may also be sufficient here.