Phonak Roger On Sphere remote mic adds neural speech separation

Phonak, a Sonova brand, has announced Roger On Sphere, a wireless remote microphone with always-on deep neural network processing. The network separates speech from background noise in real time. The device is scheduled to reach most markets from mid-November 2026.

Phonak describes it as the first remote microphone to separate speech from noise with a neural network. The announcement builds on the Roger platform, which Phonak introduced in 2013 as a proprietary 2.4 GHz digital wireless system.

How the signal is processed

The processing is called RogerSense OS and has four stages: MultiBeam directionality, always-on Sphere denoising, natural voice processing and adaptive gain. Three physical microphones feed the array, and three more virtual microphones are interpolated from them. That allows up to 12 directional beams. In table mode the device uses up to six beams around the table and picks the one with the best signal-to-noise ratio.

The selected signal then goes through the onboard network. Phonak specifies up to eight hours of continuous noise suppression. Natural voice processing uses a self-adaptive amplification strategy meant to keep the speech envelope intact, and adaptive gain follows the noise floor around the user.

An accelerometer detects whether the microphone is worn vertically, lying flat on a table or held at an angle, and chooses the matching mode. A new own-voice comfort setting in the myRogerMic app reduces the backward-facing pickup in table mode. In headset use, the network and gating also clean up outgoing speech during online calls.

Test figures and range

In Phonak's table-mode test the microphone stood 0.75 m from a talking head, with five loudspeakers playing diffuse café noise. Speech was fixed at 65 dB SPL and the input signal-to-noise ratio ran from +5 dB to -20 dB. Roger On Sphere improved the ratio by up to 14.2 dB over an unprocessed omnidirectional reference. Against Roger On 3, the extra gain reached 3 dB at an input ratio of -15 dB, which Phonak equates to about half the output noise power.

A second test placed the unit 20 cm from the talking head, with a KEMAR manikin wearing Audéo EON hearing aids set for a standard N3 hearing loss. Using the HASPIv2 intelligibility model, Phonak calculated a 12.9 dB improvement in the ratio needed for 50 percent intelligibility, compared with the hearing aids alone. These are technical measurements and modeled predictions, and clinical studies are still underway.

Transport latency is below 20 ms. Redesigned hardware and antennas give up to five times the range of Roger On 3 in free-field, clear line-of-sight conditions. The MultiTalker Network now takes up to 10 Roger On Sphere devices mixing table and presenter modes.

The unit works with hearing aids and cochlear implant systems through integrated Roger receivers, Roger X and Roger Neckloop, and with Roger Focus II ear-level receivers. "Roger On Sphere brings AI-driven speech separation to a remote microphone," says Alain Kertész, Senior Director Roger.

Source: audioXpress