Skip to content
BKtensorPublic

About

Set of πŸ“ with πŸ”— to help those building Voice AI agents πŸŽ™οΈπŸ€–

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Β 
Β 

Latest commit

Β 

History

68 Commits

Folders and files

NameName
Last commit message
Last commit date
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

Voice AI: a curated learning path for building real-time voice agents

A curated, developer-friendly learning path for building real-time voice AI agents, from your first STT call to scaling production telephony.

Awesome License: MIT Stars Last commit Resources PRs welcome

English Β· δΈ­ζ–‡η‰ˆζœ¬

Voice AI has moved from research demos into shipping product in under three years. The modern stack is converging around a clear pattern: a real-time transport layer (WebRTC or telephony), a streaming pipeline of speech-to-text β†’ LLM β†’ text-to-speech, and a turn-taking model that decides when the agent should speak. This list is structured to mirror that learning order: start with the foundations, pick a framework, then drill into individual components and production concerns.

Learning resources are tagged 🟒 Beginner, 🟑 Intermediate, or πŸ”΄ Advanced (blogs, podcasts, communities, and conferences in sections 17-20 are intentionally left untagged). Prefer free official docs and vendor-neutral guides; flag where authors have commercial interests.


How to use this list

Read top-to-bottom if you're brand new. The recommended path:

  1. Foundations β†’ understand the pipeline and latency budget
  2. Frameworks β†’ pick one (LiveKit Agents or Pipecat are the safest open-source bets) and ship a hello-world
  3. Components (STT, TTS, LLM, VAD, turn detection) β†’ swap pieces to learn what each layer does
  4. Transport & telephony β†’ connect to a real phone number
  5. Evaluation, production, ethics β†’ make it safe enough to ship

πŸ“˜ Companion book: Voice Agents Handbook

If you want this material in a tighter, opinionated, production-grade form, I wrote the Voice Agents Handbook: building production voice AI with LiveKit, plus appendices on choosing your stack and the LiveKit ecosystem beyond agents. Available now on Kindle (and in paperback).

The README you're reading collects the field's best free resources. The book is the curated path through them, with the patterns I've used shipping voice agents for trade people, lawyers, and immigration consultants.

Disclosure: I maintain this repo and authored the handbook. Free sample (Introduction + Chapter 1) at handbook.mahimai.ca.


Table of contents

πŸ“– Expand the 21 sections
  1. Foundational concepts and learning paths
  2. Frameworks and orchestration platforms
  3. Speech-to-text (STT / ASR)
  4. Text-to-speech (TTS)
  5. LLMs for voice and real-time AI
  6. Voice activity detection and turn-taking
  7. Audio enhancement and noise suppression
  8. WebRTC fundamentals
  9. Telephony and SIP
  10. Tutorials and hands-on projects
  11. GitHub starter repos and awesome lists
  12. Datasets and benchmarks
  13. Beginner-accessible research papers
  14. Evaluation and testing
  15. Production, deployment, and scaling
  16. Ethics, safety, and regulation
  17. Blogs and newsletters
  18. Podcasts
  19. Communities
  20. Conferences and events
  21. Hackathons and competitions

🧭 1. Foundational concepts and learning paths

Start here. These resources establish the mental model of the voice agent pipeline and the latency budget you'll fight for the rest of your career.

6 resources

🧩 2. Frameworks and orchestration platforms

The frameworks below all let you wire STT, an LLM, and TTS together. For open-source production work, LiveKit Agents and Pipecat are the two safest bets; for managed dashboards, Vapi, Retell, and Bland win on time-to-first-call.

Pick Type Best for
LiveKit Agents Open source Production, WebRTC-native
Pipecat Open source Vendor-neutral pipelines
Vapi / Retell / Bland Managed Fastest first call
OpenAI Realtime / GPT-Live / Gemini Live Realtime API Speech-to-speech
16 resources

Open-source frameworks

  • 🟒 LiveKit Agents: Voice AI Quickstart: Working assistant in <10 min via Python or TypeScript, runs on top of WebRTC.
  • 🟒 Pipecat: Quickstart: Scaffolds a Deepgram + OpenAI + Cartesia pipeline via the Pipecat CLI (uv tool install "pipecat-ai[cli]", then pipecat init quickstart); talk to it in the browser in ~5 minutes.
  • πŸ”΄ Ultravox (fixie-ai/ultravox): Open-weight speech LLM (a Whisper encoder feeding an LLM backbone; v0.7 runs on GLM-4.6, earlier releases on Llama, Gemma, and Qwen) that skips the separate ASR stage.
  • 🟑 qwen-audio-agent: Full-duplex voice runtime that drives coding agents (OpenCode, Claude Code, Codex, and other ACP backends) with barge-in, parallel background tasks, and a local wake word.
  • 🟑 TEN Framework: Open-source framework for real-time multimodal (voice, video, text) agents with pluggable ASR, LLM, and TTS extensions; licensed Apache 2.0 with additional restrictions and backed by Agora.

Managed platforms

Realtime / speech-to-speech APIs

  • 🟑 OpenAI Realtime API: Guide: Official guide to gpt-realtime-2.1 and gpt-realtime-2.1-mini (GA; configurable reasoning) over WebRTC, WebSockets, or SIP.
  • 🟑 OpenAI GPT-Live: Guide: Full-duplex voice model (gpt-live-1) that listens and speaks at once, decides when to yield or interrupt, and hands reasoning and tool calls to a backend text model; over WebRTC, WebSockets, or SIP.
  • 🟑 Google Gemini Live API: Overview: Low-latency, bidirectional voice + vision agents with barge-in and tool use, on gemini-3.8-live (plus an Extended Thinking variant).
  • 🟑 Twilio ConversationRelay: WebSocket bridge that handles STT/TTS so you focus on LLM logic; works with any LLM.

Vendor-neutral comparisons

🎧 3. Speech-to-text (STT / ASR)

Pick one streaming STT and learn it deeply before shopping around. Deepgram, AssemblyAI, and Whisper-derivatives cover most use cases. (All-in-one ASR + end-of-turn models like Deepgram Flux are covered under turn-taking.)

Pick Type Best for
Deepgram Nova-3 Commercial General-purpose, 36+ languages
AssemblyAI Universal-3.5 Pro Commercial Accuracy, diarization
Speechmatics Agent STT Commercial Agent turns: names, numbers
Soniox Commercial Multilingual + built-in translation
faster-whisper Open source Self-hosted Whisper
NVIDIA Parakeet (NeMo) Open source Top-of-leaderboard accuracy
25 resources

Commercial APIs

  • 🟒 Deepgram Nova-3: STT benchmarks: Primer on WER, latency, and cost alongside Deepgram's product reference; Nova-3 spans 36+ languages with multilingual code-switching.
  • 🟑 AssemblyAI Universal-3.5 Pro Streaming: Streaming STT walkthrough that doubles as a function-calling tutorial; Universal-3.5 Pro (July 2026) is now the flagship for streaming and async, with code-switching across 18 languages and earlier conversation turns accepted as context mid-call.
  • 🟒 OpenAI Speech-to-Text (gpt-transcribe / gpt-live-transcribe): Easiest cloud STT if you already use OpenAI; gpt-transcribe ($0.0045/min) and the streaming gpt-live-transcribe (July 2026) succeed gpt-4o-transcribe.
  • 🟒 Cartesia Ink 2: GA streaming STT with built-in eager turn detection and noise robustness, paired with Sonic TTS for a single-vendor low-latency stack.
  • 🟒 Soniox Speech-to-Text: One model spanning 60+ languages with real-time WebSocket streaming and async APIs, speaker diarization, language identification, endpoint detection, and built-in real-time speech translation (one-way or two-way).
  • 🟑 Speechmatics Melia: Single-pass multilingual STT with native code-switching across 55+ languages (batch, production preview).
  • 🟑 Speechmatics Agent STT: Streaming STT for voice agents (September 2026) on the Linden model, tuned for names, numbers, negations, and one-word confirmations; 369 ms median finalization in the open Pipecat benchmark.
  • 🟑 Gladia Solaria-3: STT tuned for noisy, multi-speaker European business audio (9.6% WER on English production calls).
  • 🟒 Gradium STT: Streaming STT with built-in semantic VAD; step messages every 80 ms carry end-of-turn probabilities so agents can decide when a speaker has finished.
  • 🟑 Meta Muse Voice Transcribe: One streaming model for ASR, diarization (20+ speakers), and endpointing across 25 languages with code-switching (September 2026); the most accurate model in the Pipecat streaming benchmark, though with slower P95 finalization.
  • 🟒 Gemini 3.5 Transcribe: Google's dedicated STT (August 2026) with a batch model and a streaming Live API model (gemini-3.5-transcribe-live, preview) across 85+ languages.

Open source

  • 🟒 openai/whisper: The original repo and the de facto starting point for any DIY ASR project.
  • 🟑 SYSTRAN/faster-whisper: CTranslate2 reimplementation up to 4Γ— faster with INT8; recommended for self-hosted Whisper.
  • πŸ”΄ NVIDIA NeMo (Parakeet / Canary): Top-of-leaderboard open ASR models with streaming inference recipes.
  • 🟑 FunASR: Industrial open-source ASR toolkit with Paraformer, SenseVoice, VAD, punctuation, diarization, streaming services, and an OpenAI-compatible API for self-hosted voice agents.
  • 🟑 SenseVoice: Multilingual speech understanding model for ASR, language identification, emotion recognition, and audio event detection, with FunASR integration and ONNX/libtorch export examples.
  • 🟑 Moonshine: On-device voice toolkit (STT, TTS, and an agent loop) with MIT models from 1 MB up to beyond Whisper Large V3 accuracy, running in Python, WASM, iOS, Android, and on Raspberry Pi.
  • πŸ”΄ NVIDIA Nemotron 3.5 ASR Streaming 0.6B: Open-weights cache-aware FastConformer streaming ASR across 40 locales, with runtime-configurable latency (80 ms to 1.1 s).
  • 🟑 Qwen3-ASR: Open-weight ASR from the Qwen team covering 52 languages with streaming inference; a strong self-hosted multilingual option.

Benchmarks and explainers

πŸ—£οΈ 4. Text-to-speech (TTS)

Latency, not raw quality, is what kills voice agents: prioritize providers offering true streaming with first-byte under 200 ms.

Pick Type Best for
ElevenLabs Commercial Quality, voice cloning
Cartesia Sonic Commercial Lowest latency for agents
Deepgram Aura-2 Commercial Pairs with Deepgram STT
Kokoro Open source Tiny, CPU-friendly
Chatterbox Open source Cloning + emotion control
17 resources

Commercial APIs

  • 🟒 ElevenLabs Docs: Industry-leading quality, voice cloning, and Agents platform in one SDK.
  • 🟒 Cartesia Sonic Quickstart: Sonic 3.6 (44 languages, dated snapshots, native turn detection), sub-90 ms first-byte latency, designed specifically for voice agents.
  • 🟒 Deepgram Aura-2: Low-latency streaming TTS (Aura-2) that pairs cleanly with Deepgram STT.
  • 🟒 OpenAI TTS (gpt-4o-mini-tts): Easiest plug-in TTS for the OpenAI stack.
  • 🟒 Soniox Text-to-Speech: Low-latency streaming TTS over WebSocket with multilingual voices; pairs with Soniox STT and translation.
  • 🟒 Gradium TTS: Streaming WebSocket TTS with 158 ms P50 time to first audio and instant voice cloning in five languages.
  • 🟒 Artificial Analysis: TTS leaderboard: ELO, price, and speed comparison covering Cartesia, ElevenLabs, Google, Inworld, Rime, Hume, and open-weights models.
  • 🟑 Best Text-to-Speech Providers in 2026 (Coval): Independent head-to-head of 14 TTS providers on latency, naturalness, and cost; note the commercial author.

Open source

  • 🟑 Chatterbox (resemble-ai/chatterbox): Resemble AI's MIT-licensed TTS (vendor-reported to beat ElevenLabs in blind preference tests); ~5 s zero-shot voice cloning, emotion-exaggeration control, and a built-in PerTh watermark. Turbo (350M) hits sub-150 ms first audio, Nano (110M) runs 3Γ— real time on CPU, and Multilingual (V3, 0.5B) covers 23+ languages.
  • 🟒 Kokoro 82M (kokoro-onnx): Maintained ONNX runtime for the tiny Apache-licensed Kokoro model that tops community ELO arenas; runs on CPU (the upstream weights have not changed since 2025).
  • 🟒 Piper (OHF-Voice/piper1-gpl): Fast local neural TTS optimized for Raspberry Pi; perfect for offline projects.
  • 🟑 Coqui TTS (idiap fork): Maintained fork of Coqui-TTS / XTTS v2; still battle-tested, though Chatterbox now leads on zero-shot cloning quality.
  • 🟑 Orpheus-TTS: Llama-3B-based emotive TTS with ~200 ms streaming and emotion tags.
  • πŸ”΄ Sesame CSM (sesame/csm-1b): Conversational, context-aware multi-speaker TTS using a Llama backbone with the Mimi codec; runs natively in HF Transformers.
  • 🟑 Qwen3-TTS: Apache-2.0 open-weight TTS from the Qwen team with streaming output across 10 languages.

Streaming and ethics

🧠 5. LLMs for voice and real-time AI

A voice agent's perceived intelligence is bounded by how fast the LLM streams its first token. Sub-300 ms TTFT changes the conversation feel entirely.

14 resources

Low-latency inference

  • 🟒 Groq: LPU-based inference cloud; among the lowest time to first token on mainstream open models in 2026 benchmarks.
  • 🟒 Cerebras Inference: Wafer-scale chip inference with very high throughput on Llama models.
  • 🟒 SambaCloud (SambaNova): Reconfigurable Dataflow inference; stable throughput at low latency.
  • πŸ”΄ PhoneLLM Alpha 1 (Pipecat): Open-weight (BSD-2) fine-tune of Nemotron 3 Nano 30B-A3B (3.5B active parameters) built for phone agents, with P95 time to first token under 100 ms on one B200; note the commercial author.

Speech-to-speech models

  • 🟑 OpenAI Realtime API guide: Flagship S2S product with WebRTC/WebSocket transport (gpt-realtime-2.1, GA).
  • 🟑 Google Gemini Live: Real-time multimodal voice/video with barge-in and broad language support, on gemini-3.8-live.
  • πŸ”΄ Moshi (kyutai-labs): Open full-duplex speech-text foundation model (~200 ms, Mimi codec). Kyutai's broader stack now includes Unmute (cascaded STT+LLM+TTS with tool use), Kyutai STT/TTS, and Hibiki (streaming translation).
  • 🟑 Speech-to-Speech Models in 2026: Three Architectural Bets (Krzysztof Sopyla): Vendor-neutral comparison of full-duplex (Moshi), near-duplex multimodal (Qwen-Omni), and cascade approaches, with FullDuplexBench numbers and tradeoffs.
  • 🟒 Artificial Analysis: Speech-to-Speech: Independent leaderboard for speech-to-speech models on reasoning (Big Bench Audio), conversational dynamics (Full-Duplex-Bench), latency, and price, plus a blind arena.

Voice-specific prompting and tools

πŸ”€ 6. Voice activity detection and turn-taking

Pure VAD is no longer enough: modern agents combine acoustic VAD with a small semantic model that predicts end-of-utterance from words and prosody.

13 resources
  • 🟒 Silero VAD: MIT-licensed pre-trained VAD; <1 ms per chunk on CPU. The de facto VAD inside LiveKit and Pipecat.
  • 🟑 LiveKit Turn Detector v1.0: Audio-native end-of-turn model (fused semantic + acoustic, no transcript) across 14 languages; now the LiveKit default.
  • 🟑 Deepgram Flux: All-in-one conversational STT with built-in end-of-turn detection (English median EOT <300 ms; Flux Multilingual adds 10 languages with mid-call switching), integrated with Deepgram's Voice Agent API; collapses STT and turn detection into a single model.
  • 🟑 Pipecat Smart Turn v3.2: Open BSD-2 audio-native turn detection covering 23 languages; the 8 MB int8 CPU build runs in about 10 ms on fast CPUs, with better handling of background noise and short replies.
  • 🟑 pipecat-ai/smart-turn: Repo with model code, training scripts, and integration examples (~8M params, Whisper-Tiny base).
  • 🟑 Krisp Turn-Taking v2 (VIVA SDK): Commercial audio-only turn-taking model with interruption prediction that separates barge-ins from backchannels; works with any STT/LLM/TTS stack.
  • 🟒 The Complete Guide to AI Turn-Taking (Tavus): Reader-friendly overview of why pure VAD fails in real conversations.
  • 🟑 Tackling Turn Detection in Voice AI (Notch): Engineer-first walkthrough combining VAD probability, volume, and TTS markers.
  • 🟑 What Is Endpointing in Voice AI? (Cekura): Explainer on the three-signal endpointing stack with a testing angle; note the commercial author.
  • 🟑 Evaluating End-of-Turn Detection Models (Deepgram): Methodology plus a head-to-head of Flux, Pipecat Smart Turn, and LiveKit EOU; note the commercial author.
  • 🟑 eot-bench (LiveKit): Open, reproducible end-of-turn benchmark that scores models at real pauses under latency and false-cutoff budgets, with an Apache-2.0 dataset of real human-to-agent turns in 14 languages; note the commercial author.
  • 🟑 Semantic VAD (Gradium): Explainer on multi-horizon end-of-utterance prediction, with streaming inactivity probabilities every 80 ms and guidance on tuning the latency vs interruption tradeoff; note the commercial author.
  • 🟒 ai-coustics VAD: VAD bundled with real-time speech enhancement, noise suppression, and voice isolation in a single audio preprocessing SDK; useful when you want cleanup and turn-taking signals from the same component.

πŸ”Š 7. Audio enhancement and noise suppression

The audio reaching your VAD and STT is often noisy, reverberant, or mixed with background voices. Cleaning the signal before the rest of the pipeline is frequently the difference between an agent that ships and one that frustrates users in real-world conditions (cars, cafΓ©s, call centres). In 2026 every major voice-AI vendor ships a deep-learning suppressor on top of WebRTC's classic noise-suppression chain.

6 resources
  • 🟒 ai-coustics: Real-time speech enhancement SDK covering noise cancellation, voice isolation, and VAD; on-device and cloud deployment. See the docs and developer platform.
  • 🟒 Krisp VIVA SDK: Commercial-grade real-time noise and background-voice cancellation, with VAD and turn-taking in the same SDK; the de facto standard for voice comms. LiveKit's background voice cancellation and Pipecat Cloud both build on Krisp.
  • 🟑 DPDFNet (ceva-ip/DPDFNet): Apache-2.0 real-time speech enhancement that extends DeepFilterNet2 with dual-path RNN blocks; 2.3M to 3.6M-parameter models for 8, 16, and 48 kHz audio with ONNX and TFLite exports; note the commercial author.
  • 🟑 GTCRN (Xiaobin-Rong/gtcrn): Ultra-light speech enhancement model (48.2K parameters, 33 MMACs/s, ICASSP 2024, MIT); a tiny, well-understood baseline for constrained devices.
  • 🟒 Koala Noise Suppression (Picovoice): On-device, cross-platform voice isolation with self-serve access (browser, mobile, desktop, Raspberry Pi).
  • 🟑 Noise Suppression Guide 2026 (Picovoice): Algorithms, intelligibility metrics (SII / STI / STOI), and implementation tradeoffs; note the commercial author.

🌐 8. WebRTC fundamentals

WebRTC is the default transport for voice agents that don't run over the phone network. Understanding ICE, STUN, TURN, and SFU architecture is non-negotiable for production work.

10 resources

☎️ 9. Telephony and SIP

The phone network has its own physics. Once you know which SIP trunk provider to point at LiveKit or Pipecat, you can ship.

Pick Type Best for
Twilio Commercial Default, largest ecosystem
Telnyx / Plivo Commercial Cost-effective SIP trunks
SignalWire Commercial FreeSWITCH-based, programmable
LiveKit SIP / Pipecat Framework Wire trunks into your agent
11 resources

πŸ› οΈ 10. Tutorials and hands-on projects

Pick one tutorial and finish it before starting another. Voice AI is unforgiving of half-built pipelines.

8 resources

πŸ“¦ 11. GitHub starter repos and awesome lists

Clone these instead of writing boilerplate from scratch.

7 resources
  • πŸŸ’β†’πŸ”΄ livekit/agents: The flagship open-source Python/Node framework for production voice agents (tip: pair it with the LiveKit Docs MCP server and Agent Skill for AI-assisted builds).
  • πŸŸ’β†’πŸ”΄ pipecat-ai/pipecat: Vendor-neutral framework with 80+ STT/LLM/TTS service integrations.
  • 🟒 livekit-examples/agent-starter-python: Production-ready starter with Dockerfile, CI-run conversation simulations, a multilingual turn detector, and core plugins.
  • 🟒 livekit-examples (org): Official collection of LiveKit Python/React/Swift/Android starters.
  • πŸŸ’β†’πŸŸ‘ pipecat-ai/pipecat-examples: Sample apps for push-to-talk, websocket, telephony, and multimodal use cases.
  • 🟒 elevenlabs/elevenlabs-examples: Runnable Next.js and Python examples for TTS, STT, and real-time agents.
  • 🟒 wildminder/awesome-ai-voice: Actively maintained 2026 list of open-source TTS, voice-cloning, and audio/music-generation models.

πŸ—‚οΈ 12. Datasets and benchmarks

You'll rarely train from scratch, but knowing which dataset a model was trained on explains its accents, languages, and failure modes.

8 resources
  • 🟒 LibriSpeech ASR Corpus: ~1,000 hours of English audiobooks; nearly every ASR paper benchmarks against it.
  • 🟒 Mozilla Common Voice: Crowdsourced multilingual dataset (295 languages in v27.0, CC0); the easiest legal way to fine-tune ASR.
  • 🟒 Common Voice on Mozilla Data Collective: The official home of Common Voice downloads since October 2025 (Scripted Speech 27.0 and Spontaneous Speech 5.0), via browser, API, or a Python SDK; free account required.
  • 🟒 LJSpeech Dataset: ~24 hours of single-speaker English audio; baseline corpus for Tacotron 2 and VITS.
  • 🟑 VCTK Corpus: ~110 English speakers with diverse accents; widely used for multi-speaker TTS.
  • 🟑 VoxCeleb (Oxford VGG): Million-utterance "in the wild" dataset for speaker identification and verification.
  • 🟑 TurnBench: Hand-labeled benchmark for end-of-turn and interruption detection (August 2026), with a 30 h two-channel test set (154 conversations, 106 speakers) and a 104 h training set, scored across 14 systems; the data is non-commercial.
  • 🟒 Coval Voice AI Benchmarks: Continuously updated, reproducible STT, TTS, and speech-to-speech latency and accuracy benchmarks on call-like audio across about 40 providers, with open code; note the commercial author.

πŸ“„ 13. Beginner-accessible research papers

These are the landmark papers behind the models you'll actually use. Read the Whisper and Common Voice papers first: they're unusually approachable.

10 resources

βœ… 14. Evaluation and testing

You can't ship what you can't measure. Voice-agent evaluation is fundamentally probabilistic: a single transcript can pass and fail across runs, so simulation and statistics matter more than fixed test cases.

13 resources
  • 🟒 Coval: Voice AI Testing Platform: Defines the core voice-agent metrics: TTFB, WER, resolution rate, simulated accents, and interruptions.
  • 🟒 Coval: How to Evaluate Voice Agents (Practical Guide): One of the most cited 2025 guides on probabilistic vs deterministic evaluation.
  • 🟒 Cekura: Metrics Overview: Predefined metrics, instruction-following checks, and simulation framework.
  • 🟑 Cekura: Performance Testing for Voice Agents: Practical 2025 guide on multi-turn simulation and edge-case generation.
  • 🟑 Hamming AI: Production-focused QA platform with simulation, load testing, and 50+ metrics.
  • 🟑 Hamming: Voice Agent Evaluation Metrics Guide: Reference of latency percentiles, WER, MOS-style quality, and task completion with formulas.
  • 🟑 LiveKit: Understand and Improve Agent Latency: Per-turn latency metrics (e2e, LLM TTFT, TTS TTFB) and where to optimize.
  • 🟒 Twilio: How Do You Know if Your Voice AI Agents Are Working?: 2025 guide arguing for business-outcome metrics over raw WER/latency; note the commercial author.
  • 🟑 Future AGI: Open-source platform to simulate, evaluate, trace, guardrail, and optimize voice and AI agent apps in one feedback loop, with persona-driven simulation (through its Simulate SDK) and 50+ eval metrics.
  • 🟑 Roark: Voice-AI QA and observability (YC W25) that turns failed production calls into replayable regression tests.
  • 🟑 Cekura for Agents (MCP server): MCP server that lets coding agents (Claude Code, Cursor, Codex) trigger and schedule voice-agent test runs.
  • 🟑 EVA (ServiceNow): Open-source bot-to-bot framework that scores cascaded and speech-to-speech voice agents on accuracy and experience across 213 enterprise scenarios; none of the 12 systems tested clears 0.5 on both; note the commercial author.
  • 🟑 LiveKit Agent Simulations: Runs an agent against LLM-driven simulated users that follow a scenario, in text mode on every commit or audio mode before release, and returns a pass or fail verdict with the transcript; in beta on LiveKit Cloud, note the commercial author.

πŸš€ 15. Production, deployment, and scaling

Real production voice infrastructure is the hardest unsolved problem in this space. Read these before quoting anyone a per-minute price.

10 resources

βš–οΈ 16. Ethics, safety, and regulation

If you're shipping a voice agent in 2026, disclosure and consent are no longer optional. EU AI Act Article 50 has applied since 2 August 2026, and the FCC treats AI voices as artificial under the TCPA.

11 resources

πŸ“° 17. Blogs and newsletters

Subscribe to two or three to stay current: the field moves quickly.

9 resources

πŸŽ™οΈ 18. Podcasts

4 resources

πŸ’¬ 19. Communities

9 resources

πŸ“… 20. Conferences and events

6 resources
  • AI Engineer World's Fair: Biggest AI-engineering conference; the Voice track has hosted major launches from ElevenLabs, Vapi, LiveKit, Pipecat, and Cartesia. The 2026 edition ran 29 June - 2 July 2026 at Moscone West, San Francisco; session recordings post free to the AI Engineer YouTube channel.
  • AI Engineer YouTube channel: All World's Fair and Summit talks are posted free; the best library of recent voice-AI talks.
  • AI Engineer Summit Online 2025: Voice playlist: Curated playlist including voice-track sessions from leading labs.
  • AIEWF 2026 Trends (Latent Space): Written recap of the five trends that defined the 2026 World's Fair (29 June - 2 July, San Francisco).
  • AGENTIC AI Summit (Modev, formerly VOICE & AI): Long-running Modev voice conference, now broadened to agentic AI; a one-day event on 7 October 2026 at the Carahsoft Conference Center, Reston, Virginia.
  • Interspeech 2026: Top academic speech-science conference; intimidating but worth knowing, since most landmark papers debut here. Sydney, Australia, 27 September - 1 October 2026; Interspeech 2027 moves to SΓ£o Paulo, Brazil (29 August - 2 September 2027).

πŸ† 21. Hackathons and competitions

3 resources
  • 🟒 ElevenHacks: A season of 11 weekly themed hackathons (spring 2026, $240K+ in prizes); the archive of briefs and winning projects is a good source of voice-agent project ideas.
  • 🟒 lablab.ai AI Hackathons: Continuous calendar of short online hackathons frequently sponsored by voice-AI vendors.
  • 🟒 Devpost: Voice AI Hackathons: Centralized search for active voice-AI hackathons; the best way to find what's open right now.

Suggested learning path

  1. Week 1: Foundations: Read the LiveKit pipeline post and Voice AI Illustrated Primer (sections 1, 8).
  2. Week 2: First agent: Finish the LiveKit or Pipecat quickstart end-to-end (sections 2, 10).
  3. Week 3: Components: Swap STT, TTS, and LLM providers; benchmark latency (sections 3, 4, 5).
  4. Week 4: Turn-taking, audio cleanup & telephony: Add Silero VAD, a turn detector, and a speech-enhancement pass; connect a SIP trunk (sections 6, 7, 9).
  5. Week 5: Production: Add evaluation, observability, and read the FCC/EU AI Act material (sections 14, 15, 16).
  6. Ongoing: Subscribe to two newsletters, follow Voice AI Space, and join the Voice AI community on LinkedIn group (sections 17, 18, 19).

Contributing

Pull requests welcome. Resources must be active in the last 12 months, accessible to developers, and vendor-neutral or clearly labeled when authored by a commercial party. Open an issue to suggest additions or removals. See CONTRIBUTING.md for the full contribution guide.

⭐ Stargazers and contributors

Star History Chart

Contributors

πŸ“œ License

MIT. Fork it, ship it.

About

Set of πŸ“ with πŸ”— to help those building Voice AI agents πŸŽ™οΈπŸ€–

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages