hermes - 💡(How to fix) Fix Weixin(WeChat) voice messages in Silk format skip STT — russian transcription broken

Problem

WeChat sends voice messages in Silk (proprietary Tencent codec). The Weixin platform in Hermes has two issues:

1. _extract_text() — returns voice_item.text from WeChat directly, which is transcribed through WeChat's Chinese pipeline. For Russian (and likely other non-Chinese languages) the transcription is garbage. The configured STT (mlx-whisper) is never called because text propagation short-circuits the media pipeline.

2. _download_voice() — has an early if voice_item.get("text"): return None, so if WeChat returned any transcription (even bad), the .silk file is never downloaded. Even when downloaded, the .silk raw bytes are returned — but Hermes gateway/STT expects WAV/MP3.

Environment

• Hermes Agent v0.14.0 (2026.5.16)
• macOS (Apple Silicon)
• STT: mlx-community/whisper-small-mlx via mlx-whisper
• WeChat (Weixin) bot account

Fix (local, not yet PR)

1. _extract_text() — ignore voice_item.text

for item in item_list:
    if item.get("type") == ITEM_VOICE:
        return ""        # ← always, STT must handle it

2. _download_voice() — always download, convert Silk→WAV

Removed the early return. After downloading the .silk blob, convert via pysilk + Python wave:

import pysilk, io, wave
with open(silk_path, "rb") as f:
    pcm_buf = io.BytesIO()
    pysilk.decode(f, pcm_buf, 24000)
    pcm_data = pcm_buf.getvalue()
with wave.open(wav_path, "wb") as w:
    w.setnchannels(1)
    w.setsampwidth(2)
    w.setframerate(24000)
    w.writeframes(pcm_data)

The .wav is then picked up by the standard gateway STT pipeline.

Dependencies

pip install silk-python (module name pysilk — Cython-backed, correct). NOT pip install pysilk (that's an empty stub 0.0.1).

Files

gateway/platforms/weixin.py — _extract_text() (~line 1000) and _download_voice() (~line 1529).