AGENT SKILL

vibevoice-elevenlabs-alternative

Microsoft open-source speech synthesis model, local and real-time, a stand-in for ElevenLabs. MIT. For clients that cannot send audio to a third party, or when the per-minute cost eats the margin.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 18
Forks
6

01 前言 Front Matter

name
vibevoice-elevenlabs-alternative
description
Microsoft open-source speech synthesis model, local and real-time, a stand-in for ElevenLabs. MIT. For clients that cannot send audio to a third party, or when the per-minute cost eats the margin.

02 资源文件 1

03 SKILL.md 正文

VibeVoice — Microsoft's Open-Source Voice Synthesis

Open-source TTS from Microsoft Research — generates voice locally, no per-minute API costs, no data sent to a third party. Complements (not replaces) the brain's existing speech skill which uses the OpenAI Audio API.

When to use

  • Client needs voice generation but cannot send transcripts to OpenAI / ElevenLabs (regulated, PHI, IP-sensitive)
  • ElevenLabs per-minute pricing is hurting unit economics on a high-volume use case
  • Want zero-API-cost voice for prototypes, demos, internal tooling
  • Need offline / air-gapped voice synthesis
  • Pair with the brain's transcribe skill for full local STT + TTS loops (transcribe handles audio → text, VibeVoice handles text → audio)

When NOT to use

  • Need the absolute best voice quality currently available — ElevenLabs is still ahead on naturalism for many languages
  • Low-latency streaming voice in a live agent loop — verify VibeVoice's streaming support before promising
  • Voice cloning of a specific person — quality / licensing depends; check current docs
  • Production deliverable where the client needs Microsoft-level commercial support — it's an OSS research project, not a hosted SaaS

Source of truth

  • Repository: github.com/microsoft/VibeVoice (verify current URL — Microsoft hosts many OSS projects)
  • License: MIT (verify against current state)
  • Org: Microsoft Research
  • Self-hosted: yes, run locally (GPU recommended for real-time)

VibeVoice vs ElevenLabs vs the brain's speech skill

Scenario Recommended
Production app, best voice quality, willing to pay ElevenLabs
OpenAI-managed voices via existing API key brain's speech skill (existing)
Regulated industry, no third-party data flow VibeVoice
Bulk generation where cost is the constraint VibeVoice
Air-gapped / on-prem deployment VibeVoice
Voice cloning of arbitrary speakers Verify each tool's current capability + license
Real-time streaming in a live agent loop Check VibeVoice streaming support before deciding

Quick start (rough — verify against current README)

# Clone the repo
git clone https://github.com/microsoft/VibeVoice
cd VibeVoice

# Follow the install instructions — typically a Python env + model weights download
# Run inference locally (CLI or Python API)

GPU is strongly recommended for real-time use; CPU is fine for batch / offline generation.

Recommendation pattern (proposals)

"For client engagements where voice output is needed but ElevenLabs cost or data residency is a blocker, recommend VibeVoice as the self-hosted alternative. Quality is good enough for most internal-tool, prototype, and accessibility use cases, and it pairs cleanly with our existing transcribe skill for full local STT + TTS loops. Keep ElevenLabs in the rec stack for premium-quality, hosted-fine use cases."

Limits to disclose to a client

  • Quality gap: ElevenLabs is still the quality leader in 2026 for many languages; VibeVoice is "good enough" for many use cases, not "best in class"
  • Compute: realtime requires GPU; CPU is batch-only
  • No vendor SLA: it's an OSS research project, not a commercial service
  • Language coverage: verify the languages the client needs are supported at the quality bar required

Related brain assets

  • speech — existing skill, OpenAI Audio API; VibeVoice is the local alternative
  • transcribe — existing skill, STT side of the loop; pairs with VibeVoice for end-to-end local audio
  • sora — when video + voice are needed together
  • AI Engineer agent — for evaluating local model deployment + GPU sizing for a specific arm

04 同仓库更多 Skills

CarlosCaPe
CarlosCaPe/octorato

yeet

Use only when the user explicitly asks to stage, commit, push, and open a GitHub pull request in one flow using the GitHub CLI (`gh`).

★ 18
CarlosCaPe
CarlosCaPe/octorato

wrangler

Cloudflare Workers CLI for deploying, developing, and managing Workers, KV, R2, D1, Vectorize, Hyperdrive, Workers AI, Containers, Queues, Workflows, Pipelines, and Secrets Store. Load before running wrangler commands to ensure correct syntax and best practices. Biases towards retrieval from Cloudflare docs over pre-trained knowledge.

★ 18
CarlosCaPe
CarlosCaPe/octorato

workspace-skill-discovery

Discover and include ALL skills across the workspace — not just global ones. Use at the start of every session or when the user asks about available skills. Ensures project-level skills (under .claude/skills/ or skills/ in any repo) are never missed.

★ 18
CarlosCaPe
CarlosCaPe/octorato

workers-best-practices

Reviews and authors Cloudflare Workers code against production best practices. Load when writing new Workers, reviewing Worker code, configuring wrangler.jsonc, or checking for common Workers anti-patterns (streaming, floating promises, global state, secrets, bindings, observability). Biases towards retrieval from Cloudflare docs over pre-trained knowledge.

★ 18
CarlosCaPe
CarlosCaPe/octorato

winui-app

Create, build and design WinUI 3 desktop apps with C# and the Windows App SDK following Microsoft's official guidance and CommunityToolkit. Covers XAML, navigation, theming, accessibility and deployment.

★ 18
CarlosCaPe
CarlosCaPe/octorato

windows-brain-script-portability

Scripts in ~/.claude/scripts/ that run on a Windows brain: use the python3.cmd shim and not bare python3, and reconfigure stdout to UTF-8 if you print symbols. Use it when writing a brain script or triaging a Windows-only crash.

★ 18