[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EWB7CBJE7SS3FN8V3XM8Y0":3,"skill-same:01M7EWB7CBJE7SS3FN8V3XM8Y0":45},{"skillNo":4,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M7EWB7CBJE7SS3FN8V3XM8Y0","ams-voice-ai-pro",null,"TanNhatCMS\u002Fjarvis","TanNhatCMS","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F113796420?v=4","01M753MV17KG5CD3152G3WQEQY","KietNT","01M7EKEDYVYQ6PWVJJMZ911VMX",0,"MIT","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fams-voice-ai-pro\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: voice-ai-pro\ndescription: Build voice features: speech-to-text, text-to-speech, streaming transcription, voice agents, and telephony pipelines. Use when adding speech interfaces to apps.\ncategory: development\n---\n\n# Voice AI Pro\n\nA builder's guide to speech technology: speech-to-text (STT), text-to-speech (TTS), real-time voice agents, and telephony. Covers the pipeline architecture, latency budgets, and the UX details that separate usable voice products from demos.\n\n## Overview\n\nA voice pipeline is a chain: **capture → voice activity detection → STT → intelligence (LLM\u002Fdialog) → TTS → playback**. End-to-end latency is the product metric that matters most — every 500ms of delay makes conversation feel broken. The architecture trick is streaming and overlap: start transcribing before the user finishes, start synthesizing before the LLM finishes, play audio while generating the rest.\n\n## When to use\n\n- Transcribing meetings, calls, or audio files (batch or streaming).\n- Adding TTS voiceovers to content or apps.\n- Building conversational voice agents (support lines, IVR replacement).\n- Wake-word \u002F push-to-talk interfaces.\n- Speaker diarization (\"who said what\") for multi-speaker audio.\n\n## Core concepts\n\n- **STT (speech-to-text).** Batch models for files (highest accuracy); streaming models for live audio (partial results, lower latency). Word-level timestamps enable captions and search.\n- **VAD (voice activity detection).** Detects speech vs silence\u002Fnoise. Drives endpointing (when the user is \"done\") and saves compute by skipping silence.\n- **Endpointing.** Deciding the user finished speaking. Too aggressive = interruptions; too lax = awkward pauses. Tune silence thresholds per use case (300–800ms typical).\n- **TTS.** Neural voices, near-human quality. Control via SSML (pauses, emphasis, pronunciation) or plain text. Voice cloning exists — treat as consent-gated.\n- **Barge-in.** Letting the user interrupt the agent mid-speech. Requires echo cancellation and fast pipeline teardown; without it, voice agents feel like bad IVR.\n- **Diarization.** Assigning transcript segments to speakers. Accuracy drops with overlapping speech and similar voices.\n- **Telephony.** SIP trunks \u002F CPaaS providers bridge phone calls to your pipeline (typically via WebSocket media streams, 8kHz μ-law audio).\n- **Duplex architecture.** Full-duplex (user and agent can speak simultaneously, like a phone call) vs half-duplex (walkie-talkie turns). Full-duplex needs barge-in + echo cancellation.\n\n## Practical workflow\n\n**1. Choose the pipeline shape.**\n- Transcription feature → batch STT API, store transcript + timestamps.\n- Live captions → streaming STT with partial results.\n- Voice agent → streaming STT + LLM + streaming TTS, all overlapped.\n\n**2. Set the latency budget.** Work backwards from a target (e.g., \u003C1.2s user-stops-speaking to agent-starts-speaking): VAD\u002Fendpointing ~300ms, STT final ~200ms, LLM first token ~400ms, TTS first audio ~300ms. Measure each stage separately.\n\n**3. Stream everything.**\n```\nmic → VAD → streaming STT (partials) → on endpoint: final transcript\nfinal transcript → LLM (stream tokens) → sentence buffer → TTS (stream audio) → playback\n```\nStart TTS on the first complete sentence, not the full response. Play while generating.\n\n**4. Handle the messy middle.** Partial transcripts will revise themselves — display them as provisional. Handle filler (\"um\"), restarts, and background noise with VAD tuning and noise suppression.\n\n**5. Telephony integration.** Receive call via SIP\u002FWebSocket, decode μ-law to PCM, run the pipeline, encode back. Test with real phone audio (8kHz, compressed) — lab-mic quality lies.\n\n**6. Consent and disclosure.** Inform callers they're speaking to an AI; get explicit consent for recording where law requires (two-party consent jurisdictions).\n\n## Common pitfalls\n\n- **Latency death by sequential design.** Waiting for full STT → full LLM → full TTS before playing anything yields 4–6s pauses. Overlap stages.\n- **No barge-in.** Users will talk over the agent; without interruption handling they must wait through every prompt.\n- **Endpointing mistuned.** Cutting users off mid-thought (too fast) or long dead air (too slow) — tune per population, and use semantic endpointing (LLM judges completeness) for complex queries.\n- **Echo\u002Ffeedback.** Speaker audio re-entering the mic creates loops. Use echo cancellation (AEC) on devices, headsets in testing.\n- **Voice cloning without consent.** Cloning a real person's voice — especially without explicit permission — is legally radioactive and banned by most TTS providers' terms. Use licensed\u002Fstock voices.\n- **Ignoring audio quality.** 8kHz phone audio, car noise, accents — test on the actual channel, not clean studio recordings. Accuracy numbers from benchmarks won't match your users.\n- **No fallback for STT failures.** Silence, unintelligible audio, or API errors need graceful handling (\"I didn't catch that — could you repeat?\") not crashes.\n- **Cost per minute.** Streaming STT + LLM + TTS compounds; a 10-minute call can cost dollars. Monitor, set session caps, and consider smaller models for the dialog layer.\n","# Voice AI Pro\n\nA builder's guide to speech technology: speech-to-text (STT), text-to-speech (TTS), real-time voice agents, and telephony. Covers the pipeline architecture, latency budgets, and the UX details that separate usable voice products from demos.\n\n## Overview\n\nA voice pipeline is a chain: **capture → voice activity detection → STT → intelligence (LLM\u002Fdialog) → TTS → playback**. End-to-end latency is the product metric that matters most — every 500ms of delay makes conversation feel broken. The architecture trick is streaming and overlap: start transcribing before the user finishes, start synthesizing before the LLM finishes, play audio while generating the rest.\n\n## When to use\n\n- Transcribing meetings, calls, or audio files (batch or streaming).\n- Adding TTS voiceovers to content or apps.\n- Building conversational voice agents (support lines, IVR replacement).\n- Wake-word \u002F push-to-talk interfaces.\n- Speaker diarization (\"who said what\") for multi-speaker audio.\n\n## Core concepts\n\n- **STT (speech-to-text).** Batch models for files (highest accuracy); streaming models for live audio (partial results, lower latency). Word-level timestamps enable captions and search.\n- **VAD (voice activity detection).** Detects speech vs silence\u002Fnoise. Drives endpointing (when the user is \"done\") and saves compute by skipping silence.\n- **Endpointing.** Deciding the user finished speaking. Too aggressive = interruptions; too lax = awkward pauses. Tune silence thresholds per use case (300–800ms typical).\n- **TTS.** Neural voices, near-human quality. Control via SSML (pauses, emphasis, pronunciation) or plain text. Voice cloning exists — treat as consent-gated.\n- **Barge-in.** Letting the user interrupt the agent mid-speech. Requires echo cancellation and fast pipeline teardown; without it, voice agents feel like bad IVR.\n- **Diarization.** Assigning transcript segments to speakers. Accuracy drops with overlapping speech and similar voices.\n- **Telephony.** SIP trunks \u002F CPaaS providers bridge phone calls to your pipeline (typically via WebSocket media streams, 8kHz μ-law audio).\n- **Duplex architecture.** Full-duplex (user and agent can speak simultaneously, like a phone call) vs half-duplex (walkie-talkie turns). Full-duplex needs barge-in + echo cancellation.\n\n## Practical workflow\n\n**1. Choose the pipeline shape.**\n- Transcription feature → batch STT API, store transcript + timestamps.\n- Live captions → streaming STT with partial results.\n- Voice agent → streaming STT + LLM + streaming TTS, all overlapped.\n\n**2. Set the latency budget.** Work backwards from a target (e.g., \u003C1.2s user-stops-speaking to agent-starts-speaking): VAD\u002Fendpointing ~300ms, STT final ~200ms, LLM first token ~400ms, TTS first audio ~300ms. Measure each stage separately.\n\n**3. Stream everything.**\n```\nmic → VAD → streaming STT (partials) → on endpoint: final transcript\nfinal transcript → LLM (stream tokens) → sentence buffer → TTS (stream audio) → playback\n```\nStart TTS on the first complete sentence, not the full response. Play while generating.\n\n**4. Handle the messy middle.** Partial transcripts will revise themselves — display them as provisional. Handle filler (\"um\"), restarts, and background noise with VAD tuning and noise suppression.\n\n**5. Telephony integration.** Receive call via SIP\u002FWebSocket, decode μ-law to PCM, run the pipeline, encode back. Test with real phone audio (8kHz, compressed) — lab-mic quality lies.\n\n**6. Consent and disclosure.** Inform callers they're speaking to an AI; get explicit consent for recording where law requires (two-party consent jurisdictions).\n\n## Common pitfalls\n\n- **Latency death by sequential design.** Waiting for full STT → full LLM → full TTS before playing anything yields 4–6s pauses. Overlap stages.\n- **No barge-in.** Users will talk over the agent; without interruption handling they must wait through every prompt.\n- **Endpointing mistuned.** Cutting users off mid-thought (too fast) or long dead air (too slow) — tune per population, and use semantic endpointing (LLM judges completeness) for complex queries.\n- **Echo\u002Ffeedback.** Speaker audio re-entering the mic creates loops. Use echo cancellation (AEC) on devices, headsets in testing.\n- **Voice cloning without consent.** Cloning a real person's voice — especially without explicit permission — is legally radioactive and banned by most TTS providers' terms. Use licensed\u002Fstock voices.\n- **Ignoring audio quality.** 8kHz phone audio, car noise, accents — test on the actual channel, not clean studio recordings. Accuracy numbers from benchmarks won't match your users.\n- **No fallback for STT failures.** Silence, unintelligible audio, or API errors need graceful handling (\"I didn't catch that — could you repeat?\") not crashes.\n- **Cost per minute.** Streaming STT + LLM + TTS compounds; a 10-minute call can cost dollars. Monitor, set session caps, and consider smaller models for the dialog layer.\n",{},"main","af388a77c2fe5e0b4fd879cca4300a0ffa9a6579","skills\u002Fams-voice-ai-pro",1,"https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis","https:\u002F\u002Fgithub.com\u002FTanNhatCMS","https:\u002F\u002Fmrkiet.name.vn","2026-11-07 14:23:08",{"source":7,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add TanNhatCMS\u002Fjarvis --skill ams-voice-ai-pro",{"tool":37,"command":38},"bunx","bunx skills add TanNhatCMS\u002Fjarvis --skill ams-voice-ai-pro",{"tool":40,"command":41},"pnpm","pnpm dlx skills add TanNhatCMS\u002Fjarvis --skill ams-voice-ai-pro","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fams-voice-ai-pro\u002FSKILL.md",true,[],[46,53,61,68,75,82],{"skillNo":47,"name":48,"nameZh":6,"description":49,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":50,"categories":51,"keywords":52,"publishedAt":18},"01M7EWYJR0J8DARY7E59Z6Y9MV","web-scrape","Turn web pages into LLM-ready markdown via the Firecrawl API (renders JS, anti-bot handling). Use when browser_open\u002Ffetch fails on JS-heavy pages, or when you need clean markdown\u002Fstructured extraction from a site. Requires FIRECRAWL_API_KEY.","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fweb-scrape\u002FSKILL.md",[],[],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":57,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":18},"01M7EWYBN5P6F2AJG3FR24T9BP","design-taste-frontend","tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML\u002FCSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, \"AI-made\", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Not for dashboards or app screens.","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Ftaste-design\u002FLICENSE","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Ftaste-design\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":18},"01M7EWY8FQFNQ2GB7NZZWGSSDX","test-driven-development","Use when implementing any feature or bugfix, before writing implementation code","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-tdd\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":18},"01M7EWY5SHBFRH1HDHVWM9K7TV","systematic-debugging","Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-debugging\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":18},"01M7EWY3NM0Z0JAKS7HDZ24K66","requesting-code-review","Use when completing tasks, implementing major features, or before merging to verify work meets requirements","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-code-review\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":18},"01M7EWY0WRMWX4M9ZFY0WKC593","brainstorming","You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).","https:\u002F\u002Fgithub.com\u002FTanNhatCMS\u002Fjarvis\u002Fblob\u002Fmain\u002Fskills\u002Fsuperpowers-brainstorming\u002FSKILL.md",[],[]]