vrchat-fishaudio-posting

VRChat Fish Audio Hakua Posting

When to Use

Use this skill for any VRChat screenshot-to-video post where the user requires Fish Audio's cloned Hakua voice and verified social delivery.

Class-level workflow for converting a VRChat screenshot into a narrated MP4 and posting it through Hermes LM Twitterer or another approved social adapter.

User voice contract

For this user, the narration audio must be Fish Audio's cloned Hakua voice. Pass the cloned-voice reference_id explicitly; never rely on a provider default. Do not silently substitute Irodori-TTS, Edge TTS, or a system voice. If Fish Audio synthesis fails or the reference ID is absent, stop before video creation and posting.

Verified configuration:

• provider: fishaudio
• model: s2.1-pro-free
• format: mp3
• Hakua reference ID: 25d6710bcdf043a7b7441e58294431f9

Never store or print API-key values.

Procedure

1. Select and validate the VRChat image.
2. Check Fish Audio status: credentials available and reference_id_configured true.
3. Generate a short Japanese narration. A narration-text timeout may use an explicitly approved text fallback, but it must never change the audio provider or voice.
4. Use the working Fish Audio core path: plugins.fish_audio_tts.core.settings() followed by synthesize_text(text=..., voice=cfg.reference_id, model=cfg.model, output_format="mp3").
5. Verify the result reports provider=fishaudio, the expected voice ID and model, and a non-empty MP3.
6. Encode the image and MP3 as H.264/AAC MP4 with yuv420p. Ensure duration is above the destination minimum; X media requires more than 500 ms.
7. Copy the MP4 to the LM Twitterer media directory when required.
8. Post only after local checks pass. Verify structured output has posted: true, a media ID, and a URL.

Failure policy

• Missing API key: stop and report configuration status without exposing it.
• Missing or unexpected reference ID: stop; do not use a default Fish voice.
• Fish Audio API/synthesis failure: stop; do not use another TTS provider.
• Encoding failure or ambiguous post response: stop and report the evidence.

Evidence discipline

Separate voice, media, and delivery evidence. A successful Irodori or Edge render proves only that another voice worked; it does not satisfy the Fish Audio Hakua requirement. Report provider/model/reference-ID configuration status, file existence and size, duration, media ID, and URL, but no secrets.

References

• references/verified-pipeline.md contains the session-verified implementation and checks.
• fishaudio-tts contains the provider API details and wrapper caveat.
• x-video-encode-and-cjk-font contains X encoding guidance.