Audio Service Architecture

AudioService owns one input engine and keeps codec I/O, Opus encoding/decoding, and network queues independent from the chip-specific speech pipeline.

Input engines

The old AudioProcessor + WakeWord combinations have been replaced by a single AudioEngine interface. AudioInputTask reads PCM once and feeds exactly one engine.

AfeAudioEngine owns a single FD AFE instance. WakeNet and voice uplink share that instance, so enabling both no longer creates two AFE pipelines. For custom MultiNet wake words, AFE fetch output is passed to CustomWakeWord; MultiNet is not created on the smaller targets.

The AFE configuration currently uses FD_LOW_COST AEC with AEC_NLP_LEVEL_VERYAGGR. WebRTC/NSNet noise suppression is intentionally disabled because the project does not ship an NSNet model.

When wake-word audio upload is enabled, the most recent two seconds of PCM are stored in a single 64 KB PSRAM ring buffer. WakeNet and MultiNet share the same cache implementation, and the encoder reads it one Opus frame at a time. This avoids the previous per-chunk internal-SRAM allocations and temporary PCM concatenation buffer.

Input data flow

flowchart LR
    Mic[Microphone] --> Codec[AudioCodec]
    Codec --> Input[AudioInputTask]
    Input --> Engine[One AudioEngine]
    Engine --> Wake[Wake-word event]
    Engine --> PCM[16 kHz mono PCM]
    PCM --> EncodeQueue[audio_encode_queue_]
    EncodeQueue --> Opus[OpusCodecTask]
    Opus --> SendQueue[audio_send_queue_]
    SendQueue --> App[Application / network]

Wake-word detection and voice processing are independent runtime states on the same engine. On AFE targets, AEC stays active while wake-word detection is active so playback reference remains available for wake-up during device playback. It also stays active during voice processing when device AEC is requested.

Output data flow

flowchart LR
    App[Application / network] --> DecodeQueue[audio_decode_queue_]
    DecodeQueue --> Opus[OpusCodecTask]
    Opus --> PlaybackQueue[audio_playback_queue_]
    PlaybackQueue --> Output[AudioOutputTask]
    Output --> Codec[AudioCodec]
    Codec --> Speaker[Speaker]

Tasks and power management

• AudioInputTask reads codec input and feeds the selected engine.
• AudioOutputTask drains decoded PCM to the codec output.
• OpusCodecTask encodes uplink PCM and decodes downlink packets.
• AfeAudioEngine has its own AFE fetch task on S3/P4.

The audio power timer still enables and disables codec ADC/DAC channels based on activity; the engine refactor does not change that policy.