Voice Translate

Use this skill for two closely related modes:

1. 1. Chat-native mode: the user sends audio or a voice note in OpenClaw; return transcript text, translation text, and translated audio.
2. Local pipeline mode: run a deterministic file-based pipeline that writes transcript, translation, wav, and metadata artifacts.

Default to an LLM-assisted translation workflow: let the current agent produce the translation, save it to a file when using the local pipeline, or use the surrounding agent turn directly when responding in chat.

Workflow

A. Chat-native mode

Use this when an inbound message already contains an audio transcript from OpenClaw media understanding, or when the user asks you to process a voice message conversationally.

1. 1. Detect that the user sent audio or that the request is for voice translation.
2. Obtain or confirm the transcript text.
3. Translate with the current model.
4. Send the transcript text to the user.
5. Send the translated text to the user.
6. Synthesize the translated text as audio:

1. 7. Keep the output order stable: transcript first, translation second, audio last.

B. Local pipeline mode

1. 1. Confirm input/output expectations: source language, target language, output directory, and whether the run should be real or mock.
2. Choose backends:

1. 3. Run transcription.
2. If using the default llm path, read the transcript and translate it with the current model. Save the translated text to a file.
3. Run synthesis/output writing with --translation-file.
4. Inspect outputs:

1. 7. If the user wants chat updates during processing, pass notifier commands with --transcript-command, --translation-command, and --audio-command.

Preferred execution patterns

Default LLM-assisted path

Use this when the agent handling the task can translate the transcript itself.

1. 1. Run the pipeline once transcription is available, or run the full command after preparing translation.txt.
2. Save the model-produced translation to a file.
3. Invoke:

CODEBLOCK0

Read references/llm-translation-pattern.md when you need the exact orchestration pattern or a reusable translation prompt.

Mock end-to-end validation

Use this first when you need to validate the pipeline structure without model/runtime dependencies.

CODEBLOCK1

Notes:

• - mock transcription reads plain text from the input file.
• INLINECODE20 TTS writes a silent wav file.
• INLINECODE21 is still required by the current CLI shape even when using mock TTS; use any placeholder path.
• INLINECODE22 mode currently means the translation must already exist in --translation-file.

Service fallback

CODEBLOCK2

Resources

scripts/

• - run_voice_translate.py: primary entrypoint.
• INLINECODE25: thin wrapper for the default LLM-assisted path.
• INLINECODE26: pipeline modules.
• INLINECODE27: wrap stage text and forward it via a shell command.
• INLINECODE28: forward generated audio via a shell command.
• INLINECODE29, mock_audio_sender.py: local smoke-test helpers.

references/

• - Read references/runtime-notes.md for dependency/setup details, backend behavior, and integration constraints.
• Read references/llm-translation-pattern.md when the surrounding agent should perform translation with its own model.
• Read references/openclaw-chat-mode.md when implementing or following the conversational flow: receive voice, output transcript text, output translation text, then output translated audio.

Editing guidance

• - Keep SKILL.md procedural and short.
• Put environment- or backend-specific detail in references.
• Treat llm as the preferred translation path for agent-driven workflows.
• In chat-native mode, preserve the user-visible ordering: transcript text, translation text, then audio.
• Prefer OpenClaw tts for immediate conversational audio replies; prefer Piper for local wav artifacts and offline pipelines.
• If the user wants tighter OpenClaw integration, add an attachment-aware outer workflow or hook instead of rewriting ASR/TTS first.
• Preserve the current file contract unless the user asks to change it: transcript, translation, wav, metadata JSON.