This need for a new generation of ‘green’ microphones leads to embedding the capability for waking up the rest of the system as soon as a voice activity is detected, possibly once a keyword has been spotted, while enhancing the recording quality with high sensitivity in far field and near field environments.
The combination of a complete analogue-to-digital converter (ADC) with a voice processing DSP has, Dolphin asserts, reached its limits for lowering power consumption. Voice detection by a DSP in always-listening mode is proving to be unsatisfactory for IoT applications powered by small batteries or even on a smartphone. Green microphones fill the low-power gap by reducing by a ratio of at least five, the power consumption in always-listening mode.
To increase battery life far beyond any current solution, and to improve the sound recording quality, such green microphones require a new generation of ultra-low power voice triggers and fast response ADC modulators. The DSP can then be embedded in the same SoC as the sound signal buffer and perform the ADC decimation and filtering functions.
Dolphin Integration, building on its work in delta-sigma converters for audio applications, has designed the first silicon IP – mMOD67.5-ROOSTER-VD.01 – targeting green microphones which combines an ultra-low power voice activity detector (WhisperTrigger-a) with an ultra-low power ADC modulator. Using a self-adaptive design, the WhisperTrigger-a performs highly accurate voice detection, with a negligible detection latency, thus permitting efficient key-word spotting or speech-to-text recognition by the DSP, whatever the user environment.
Features of the mMOD67.5-ROOSTER-VD.01 include:
• Ultra-low power consumption
• 25 µA in Voice Detection mode
• 360 µA in full performance Record mode from 20 Hz to 20 kHz, at Fmclk = 1.536 MHz
• SNR: 67.5 dB at -26 dBFs in the [20 – 20k] bandwidth (A-weighted), at Fmclk = 1.536 MHz
• Fast wake-up time of the ADC modulator: lower than 1 msec
• Efficient detection performance in far-field context (per the MIWOK-C r1.0 benchmark):
• 27 msec of average detection latency
• 97.7% of Voice Detected as a Voice (VDV) within the first phoneme
• 3.6% of Noise Detected as a Voice (NDV)
The mMOD67.5-ROOSTER is straightforward to integrate in the microphone with no need for end-user training.
The illustrations show the architecture of the IP, and a typical applications block diagram. A discussion about the company’s WhisperTrigger voice activity detector can be seen in the following video:
Dolphin Integration: www.dolphin-ip.com