gaelic-ghost/speakswiftlyserver/speak-swiftly-voice-workflows

SpeakSwiftly Voice Workflows

Use this skill for voice selection, voice creation, and speech-generation work. The local speak_swiftly MCP surface owns live speech playback and guidance; HTTP owns voice-profile mutation and retained generation.

Start Here

• Read speak-swiftly://voices or GET /voices before creating, renaming, rerolling, deleting, or choosing a profile.
• Use GET /voices for detailed HTTP workflows; the slim MCP surface no longer exposes one-profile detail resources.
• If the user wants help designing a voice rather than executing immediately, prefer the draft_profile_voice_description, draft_profile_source_text, and draft_voice_design_instruction prompts plus the guide flow documented in MCPResources.swift.

Creation And Editing

• Use POST /voices/from-description when the user has target sound qualities and source text.
• Use POST /voices/from-audio when the user has reference audio. Provide transcript whenever the spoken words are already known.
• Use PUT /voices/{profile_name}/name for a pure rename.
• Use POST /voices/{profile_name}/reroll when the user wants the same stored name rebuilt from its original inputs.
• Use DELETE /voices/{profile_name} only after confirming the exact stored profile_name.
• The package-owned built-in defaults are swift-signal and swift-anchor when installed by upstream SpeakSwiftly bundled system-profile resources. Treat them as system voices, not user-authored example names. Normal users should list and select them.
• When a user wants broad-appeal user-authored example profiles, suggest names and voice directions such as:
• swift-lumen: luminous, clean, gentle, and polished
• swift-arc: compact, focused, modern, and precise
• swift-melody: warm, expressive, friendly, and persona-ready
• swift-foundry: grounded, maker-like, textured, and steady

Qwen Voice-Design Prompting

• Treat Qwen3-TTS VoiceDesign prompts as standalone voice designs. The model receives the current spoken text plus the current natural-language instruction, not an audio memory of a previous generated profile.
• When a user wants a change to an existing synthetic voice, rewrite the full intended voice description with both the preserved traits and the requested change. Avoid relative instructions like "same as before, but warmer" or "this voice with less edge"; they leave the model without the original target voice.
• Make instructions concrete across timbre, age or gender presentation when explicitly requested, pace, pitch range, affect, accent or dialect only when requested, breath/texture, and performance style. Prefer stable acoustic language over brand names, celebrity names, or vague vibe labels.
• Keep source text aligned with the desired voice. A reusable profile source passage should exercise the cadence, emotion, and phonetic range the user wants, because Qwen's own design-then-clone workflow starts by creating a short reference clip from a matching text and instruction pair.
• For consistency across many lines, create or reroll a profile from a strong self-contained design first, then reuse that stored profile for generation. Do not try to get long-term voice identity by sending incremental VoiceDesign instructions line by line.

Speech And Artifacts

• Use generate_speech when the user wants audible playback now.
• Use POST /speech/files when the user wants a saved retained artifact instead of immediate playback.
• Use POST /speech/batches when the user wants multiple retained artifacts generated under one voice profile.
• Pass text_profile_id only when the user explicitly wants a stored normalization profile on that request.
• HTTP and MCP speech requests get transport provenance in request_context by default; pass cwd, repo_root, or explicit request_context only when path or caller metadata needs to be more specific. Do not pass a separate source_format; SpeakSwiftly's package-owned normalizer infers text and source structure from request text plus path context.
• Pass qwen_pre_model_text_chunking only when the user explicitly wants Qwen live playback to chunk before model generation; omitted requests keep the runtime's normal single-pass live path.

Tracking

• After generate_speech, read speak-swiftly://requests/{request_id} or speak-swiftly://overview.
• After retained-file or batch requests, follow GET /requests/{request_id} first, then inspect GET /generation/jobs, GET /generation/jobs/{job_id}, GET /generation/artifacts, or GET /generation/artifacts/{artifact_id}.
• Use generation jobs and artifacts to inspect retained outputs; the older generated-file and generated-batch read tools are not carried forward.
• Use DELETE /generation/jobs/{job_id} only when the user explicitly wants one retained generation job removed.