Tag
An open-source website, voicebenchmarks.com, has been built to centralize Voice AI benchmarks for TTS, LLM, STT, and more, featuring a user-friendly interface and inviting community contributions.
The article discusses hidden sources of latency in voice-agent systems after streaming is enabled and seeks advice on which performance metrics are crucial for real-time agents.
A technical breakdown of an enterprise Voice AI architecture that cuts telephony costs by 40-60% via wholesale carriers and achieves sub-500ms latency using Deepgram, Claude/GPT-4o-mini, and ElevenLabs/Cartesia, orchestrated through n8n and Supabase.
The author built a local realtime voice stack using Parakeet STT, Qwen 2.5 7B, and Qwen3-TTS, integrated with Ollama.
A developer asks what STT APIs people use in production voice agents, comparing Deepgram, AssemblyAI, and Smallest AI Pulse, and highlighting common failure points like endpointing, latency, and barge-in.
An analysis of 10,000 voice AI calls reveals key failure points: STT error rates, first-8-second chaos, interruption handling, extended silence, tool call latency, LLM failures, and broken escalation — offering practical insights for agent builders.
The article discusses the challenge of separating probabilistic language understanding from deterministic execution in voice agents, recommending a confidence check for conditional phrases before triggering structured actions, as implemented with Vomo AI.
LangChain launches LangSmith tracing for voice frameworks (Pipecat, LiveKit, OpenAI Realtime, Gemini Live), enabling full audio monitoring, STT/TTS latency tracking, interruption detection, and VAD analysis with minimal code.
The Cicero project enables self-hosted, bidirectional voice conversation with AI agents, supporting local TTS/STT and integration with various agent protocols and tools like Claude Code and Hermes Agent.
audio.cpp update introduces streaming support and four ASR/STT models (Nemotron 3.5, Higgs Audio STT, VibeVoice ASR, Hviske ASR) in native C++/GGML, achieving up to 2.41x faster than Python with competitive WER and VRAM usage.
A guide demystifying voice agents using speech-to-text and text-to-speech, with four browser-based demo agents and instructions to build your own using RAG and tools.
A new speech-to-text tool claims to rival Dragon Professional on Windows, offering a competitive alternative for voice recognition.
Describes a medical speech-to-text system that runs locally on a MacBook, enabling streaming transcription without cloud dependency.
The author argues that for live voice agents, STT latency and real-time behavior are more critical than raw transcription accuracy, and proposes a different evaluation scorecard.
A developer debunks the common belief that LLM latency is the primary cause of slow voice agents, explaining that delays often stem from earlier stages like audio capture, VAD, and STT. They recommend logging specific latency metrics and testing various STT/TTS providers and orchestration frameworks to diagnose issues.
A curated, open-source learning path for building voice agents, covering from STT to production, with 190+ resources and a 5-week plan.
pibot is now fully local, using Parakeet for STT, Qwen3-tts for TTS, and Qwen 3.6 as the local multimodal LLM via llama.cpp, with Rust/mlx-c based inference engines, achieving zero Python dependencies.
Interfaze introduces a new hybrid AI model architecture that combines DNN/CNN encoders with transformers to achieve superior accuracy and cost-efficiency for deterministic tasks such as OCR, vision, and STT, compared to generalist models.
A new AI model from interfaze_ai claims to outperform leading models (sonnet 4.6, gemini 3 flash, gpt 5.4 mini) on OCR, vision, and speech-to-text tasks.
LiveKit Agents is an open-source framework for building realtime, multimodal voice agents that can see, hear, and understand, with flexible STT/LLM/TTS integrations, job scheduling, telephony support, MCP compatibility, and a built-in test framework.