## Instructions - Choose path deliberately: ElevenLabs cloud for quality/emotion, Piper local for private/offline, XTTS for cloning. - Design voices with prompts: age + gender + texture + pace + accent, 2-3 sentences, test same 30-word script. - Require cloning consent: written release, identity check, no public-figure mimic, store consent PDF with voice id. - Disclose AI voice: verbal This is an AI voice + show notes tag, never pass robocalls as human. - Normalize text first: expand 3pm to 3 P M, $12 to twelve dollars, split 200-word chunks at sentence ends. - Set synthesis params: stability 0.5 narration / 0.3 drama, similarity 0.75 clone, 22050Hz local, MP3 128k delivery. - Pair with tts-say plugin: plugin handles playback plumbing, this skill owns voice choice, prompts, ethics. - Master loudness: -16 LUFS podcast, -14 LUFS shortform, true-peak -1.5dB, 0.8s room-tone head/tail. - Version voices: aria-v1.2 stability-50, keep sample WAV + prompt + consent hash in voices/ git folder. - Batch efficiently: off-peak cloud renders, local Piper queue overnight, cache repeated intros/outros by hash. - Handle 3 languages max per voice: separate model per language family, native-speaker check before publish. - Provide fallback: if API down, auto-switch to Piper en_US-lessac-medium with notice in show notes. ## Examples - Example 1: Podcast host: ElevenLabs aria-v1.2 stability 0.55, 30-min render, -16 LUFS, AI-voice disclosed. - Example 2: Offline kiosk: Piper en_US-lessac-medium 22kHz on Pi 5, 0.4s latency, no cloud PII leaves device. - Example 3: Consent clone: narrator release signed, 3-min clean sample, similarity 0.75, consent PDF archived. - Example 4: Refusal: public-figure impression request declined, offered original voice design instead. ## Workflows - Workflow A — Design: 1. Prompt draft 2. 30-word test 3. 3 variants 4. Listener vote 5. Lock v1.0. - Workflow B — Clone: 1. Consent sign 2. 3-min sample 3. Denoise 4. Train/XTTS ref 5. Similarity test 6. Archive. - Workflow C — Produce: 1. Normalize text 2. Chunk 200w 3. Synthesize 4. Loudness master 5. Disclosure tag. - Workflow D — Fallback: 1. Health check 2. Piper switch 3. Notice note 4. Re-render cloud later. ## Templates - Voice design prompt: `Warm female 30s, smoky texture, slow 0.9x pace, soft Irish lilt, bedtime-story calm, no fry.`. - ElevenLabs request: ```json {"voice_id": "aria-v1-2", "model_id": "eleven_multilingual_v2", "stability": 0.55, "similarity_boost": 0.75, "style": 0.3, "text": "Welcome back. Today we test three mics under forty dollars."} ``` - Piper local render: ```bash echo "Welcome back." | piper --model en_US-lessac-medium --output_file ep12.wav ffmpeg -i ep12.wav -ar 22050 -b:a 128k ep12.mp3 && loudnorm -I ep12.mp3 ``` - Disclosure line: `Voice: AI-generated (ElevenLabs aria-v1.2) with consent where cloned — full release on file.`. ## Tools - ElevenLabs — top-tier emotion/cloning, multilingual v2, starter $5/mo beats DIY for hero voices. - Piper TTS — fast local offline voices on Pi/CPU, 22050Hz, 30+ languages, zero per-char cost. - XTTS v2 / Coqui — local cloning with 3-min sample, style transfer, runs on 8GB VRAM. - Audacity / ffmpeg — denoise clone samples, loudnorm to -16/-14 LUFS, MP3 chapter encode. - tts-say plugin — SARAH playback wiring, queue, fallback switching this skill configures. - Consent form template — release text, usage scope, revocation clause stored per voice id. ## Advanced - Clone quality: 3-min 48kHz dry closet sample, RNNoise strip, -3dB peaks, no music bleed or reverb. - Tune prosody: SSML breaks 300ms sentences, emphasis on prices/names, spell acronyms with spaces. - Multilingual keep: separate XTTS checkpoint per language, 0.7 similarity cap avoids accent bleed. - Cut cost: cache intro/outro/ad-reads by text hash, reuse MP3, cloud only for new body copy. - A/B voices: same script two designs, 20-listener blind vote on warmth/clarity before series lock. - Live latency: Piper streaming chunks 15 words, 0.4s first-audio, ElevenLabs Turbo only for realtime bots. - Watermark provenance: embed AI-voice tag in ID3 + show notes URL to consent record for disputes. - Monitor drift: re-render golden 30-word script monthly, diff MOS score, repin model_id on change. - Backup voices: export ElevenLabs voice settings + Piper .onnx to voices/ with consent hash quarterly. - Test intelligibility: run 100-word number/acronym script, fix misreads via alias dictionary before series. ## Documentation - ElevenLabs docs — voices, models, stability/similarity tuning, cloning consent and pricing. - Piper README — model list, sample rates, CLI flags and Pi 5 performance table. - XTTS v2 docs — sample requirements, VRAM table, multilingual limits and fine-tune steps. - FTC/FCC robocall rules — AI-voice disclosure, consent, and telemarketing restriction summaries. - Loudness standards — podcast -16 vs shortform -14 LUFS, true-peak and delivery specs.