Add ESP32-P4 support; fix voice sequence numbers and VAD cut-offs

[haarts]

Adds support for the Waveshare ESP32-P4-WIFI6-POE-ETH and fixes two audio problems that affect every board. Three commits that can be reviewed separately:

1. Add ESP32-P4 support

• Opus: OPUS_XTENSA_LX7 is now set only for the ESP32-S3 instead of in library.json, so the P4 (RISC-V) uses the generic C code. lib/micro-opus/CMakeLists.txt registers the vendored Opus as an IDF component for ESPHome's native esp-idf toolchain.
• Toolchain: the P4 config uses the native esp-idf toolchain, because PlatformIO's can't link rev3 P4 builds at the moment (pioarduino 55.03.39 expects sections.rev3.ld.in, which its ESP-IDF 5.5.5 doesn't ship). The S3 configs keep using PlatformIO, unchanged.
• Wi-Fi-only calls are guarded; the default username uses the Ethernet MAC when there's no Wi-Fi.
• New mic_warmup option: settle time after acquiring the I2S bus. It defaults to the previous 200 ms; the P4 config uses 50 ms.
• Board config esphome/waveshare-esp32-p4-poe.yaml: IP101 Ethernet, ES8311 with the board's analog mic. ESPHome's use_microphone: true selects a PDM mic, so it stays off. Mic preamp and digital gain are runtime settings (the defaults of 72 dB in total overdrove the input).

2. Fix voice packet sequence numbers (all boards)

The voice sequence counts 10 ms frames: desktop Mumble advances it by 2 per 20 ms Opus packet and keeps it running across utterances. We advanced it by 1 and reset it to 0 at every utterance. Mumla and desktop Mumble time their jitter buffer by it and dropped most of the audio, so speech came through in fragments of a few hundred ms. pymumble and the server don't care, which is probably why it went unnoticed.

3. Improve voice detection and push-to-talk (all boards)

• The adaptive noise floor crept up under continuous speech and cut TX after about a second. It now only learns the ambient level between utterances.
• Hangover goes from 300 to 800 ms, so pauses between words don't end TX.
• Push-to-talk sends everything while the button is held, with no VAD gating and no calibration window swallowing the first word.
• Always-on sends a 200 ms pre-roll when VAD triggers.

Testing

• On hardware: Waveshare ESP32-P4-WIFI6-POE-ETH (rev v3.2), against Murmur 1.5 (official Docker image), listening in Mumla on Android.
• End to end: speech was played next to the board, recorded from the server with a pymumble client, and transcribed with Whisper. A Dutch test sentence went from about 60% word error rate to 0% in both always-on and push-to-talk.
• Builds: pre-commit run --all-files passes on each commit. The P4 config compiles at each commit, and esp32-s3-box.yaml compiles with all three.
• Not tested on S3 hardware: I don't have an S3 board. Commits 2 and 3 change shared behavior, so a quick check on a Box or Voice PE would be welcome.

I haven't added the P4 board to the CI matrix or the release/OTA manifests, because it needs the native esp-idf toolchain instead of PlatformIO. Happy to add that if you want it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NmT8zZQWZTiXj38MZhU3tT