Each spring, tech hobbyists visit Alibaba for its annual open tour. This May 10, Synced toured the company’s acoustic lab, which is dedicated to R&D for its voice assistant Tmall Genie.
The acoustic lab looks pretty ordinary, researchers hover around a bunch of Tmall Genie and microphone parts piled on a work bench. Hidden behind closed doors however is the lab’s jewel: the anechoic and reverberation rooms.
The anechoic room is less than 5 sqm in area. One enters via a thick door that shuts firmly with rubber seals. The ceiling, floor, and walls are covered with convex polyhedral baffles — acoustic wedge panels capable of absorbing all incoming sound without reflection.
If you’ve ever experienced that sound dampening ear effect on an airplane or due to water clog, this is what it’s like in the room. One’s voice here also sound much deeper than in other environments.
Zhi Hai, the lab’s sound engineer tells us, “With a threshold frequency of 100 Hz or less, environmental noise in this room can go below 20dB.”
High-sensitivity microphones in the room test speakers for response frequency, audio distortion, and overall sound effect performance, explains Hai. “The placement of the microphones and the overall design of arrays will affect audio pickup. Microphone voice input is very important, as audio recording and signal processing are keys for voice recognition.” Lab engineers optimize microphones arrays to obtain high-quality sound using cost-effective designs.
Two types of tests are performed in the reverberation room: measurement of speaker efficiency; and voice recognition with generative artificial reverberation, where engineers take natural soundscapes to the extreme to improve voice recognition accuracy.
Three noisy air purifiers and seven pieces of audio equipment are placed inside the room, as engineer Hai explains, “for simulation of real noise.” In addition to the walls’ sound absorptive materials, removable wooden boards and metal panels are configured to adjust reverberation time. Hai demonstrates by altering a baffle board’s position, “Now that I have changed sound absorption, this room will have a different reverberation time even though you may not feel it.” Long reverberation time makes the audio a lot more enjoyable, but also negatively affects clarity; while shorter reverberation time can make audio sound dry and unnatural.
Hai explains why smart speaker manufacturers use anechoic and reverberation rooms for speaker testing: “The anechoic room is used to test speakers under an ideal environment, which definitely helps with product development. Experiments performed with artificially generated noises in the reverberation room meanwhile can test voice recognition and acoustic performance from various directions, which is crucial for product improvement.”
We ask about the multiple black cables extending through the doors of both rooms. Hai explains these are server connections, so engineers to perform remote monitoring, artificial voice input, data testing, and collection and analysis at the platform without entering the room.
Before leaving we notice a tester speaking into a microphone while data rapidly refreshes on their computer screen. “This is test done on a daily basis,” an Alibaba employee tells us, “Testers work with acoustic test data, to modify and design the system for digital signal processing. All AI lab group members enjoy visiting our acoustics laboratory to discuss with us, as ‘sound’ is the key component for Tmall Genie.”
Since Alibaba expanded into cloud computing, AI, and big data, the company has deployed a number of new labs at its Hangzhou headquarter, including AI labs, a machine intelligence lab, and a cybersecurity lab.
Localization: Tingting Cao | Editor: Michael Sarazen