ElevenLabs

ElevenLabs adds multimodal inputs to ElevenAgents: images, PDFs, audio messages, contacts and locations on WhatsApp, and image and PDF uploads in the web widget, with context carried across voice, chat and messaging within one conversation.

ElevenLabs releases Eleven v4 and a low-latency Turbo variant for text-to-speech, with controllable delivery through inline tags, better multi-speaker consistency, improved voice cloning and phoneme support, and about 100ms median latency for agents.

ElevenLabs publishes an official Power for the Kiro IDE covering its speech, transcription, agents, sound effect and music APIs. Steering files load by prompt intent, which keeps agent context small compared with static skills or manually linked docs.

ElevenLabs updated its Scribe v2 transcription model with PII redaction at transcription time, correct Latin-script English inside Indic speech, a filler-free No Verbatim mode, and keyterm prompting expanded to 1,000 terms.

ElevenLabs walks through three Text to Speech integration modes, model and format choice, concurrency-bounded batching, text splitting, parameter-hash caching, 429 backoff with jitter, and time-to-first-byte benchmarking for production voice apps and agents.

ElevenLabs explains why recycling human-training playbooks produces weak conversational agents. It covers the default LLM gaps in tone and scope for speech, configuration choices, and guardrails against unpredictable behavior.

ElevenLabs recaps a workshop on enterprise agent safety. It frames controls around input, model behavior, and output, then demos guardrail configuration, pre-launch simulation testing, and PII redaction for sensitive deployments.

ElevenLabs describes how it monitors and tests its documentation agent. It uses evaluation criteria such as hallucination and resolution, full and partial conversation simulations with tool mocking, and CI/CD-automated test runs.

A walkthrough for building an inbound healthcare scheduling voice agent in ElevenAgents. It covers system prompt, guardrails, tools against a sample FHIR EHR and calendar, patient verification, escalation to staff, analysis and testing, and Twilio telephony.

ElevenLabs explains audio tags as inline performance directions for its v3 and v4 TTS models, mapping SSML equivalents and covering tag categories, combination, and punctuation effects on pacing.

ElevenLabs adds Versioning to its voice agent platform. Each version captures the full configuration, with diffs and editor attribution. Traffic splitting lets teams A/B test a new version before migrating, and the config history gives an audit trail.

Argues that production agents fail from architecture, not models: one agent carrying every responsibility has no checkpoint between understanding and acting. Proposes selective specialization, with the tradeoffs and where ROI comes from.

ElevenLabs releases Music v2 with better vocals and multilingual output, section-by-section song building and inpainting of single sections. It ships in ElevenMusic, ElevenAPI and ElevenCreative, with price cuts of up to 50% on the API.

ElevenLabs launches Realtime Translation in enterprise alpha on ElevenAPI. It translates live calls across 90+ languages with automatic detection on Eleven v4 Turbo, keeps one chosen voice per session, and lets teams trade translation quality against latency.

ElevenLabs defines conversational AI latency across the speech-to-text, LLM and text-to-speech pipeline, separating model inference latency, LLM time to first token and TTS time to first audio.

ElevenLabs details how it built a voice agent that interviewed 230+ app users in under 24 hours. Covers Gemini 2.5 Flash for logic, prompt guardrails for vague answers, simulated tests, and structured data extraction from transcripts.

ElevenLabs walks through Agent Skills as folders of markdown plus scripts and references. Only frontmatter loads by default, with deeper files pulled in on demand to limit context bloat, and it shows how to use shared skills in coding workflows.

ElevenLabs ships Music v2.5 as the default in ElevenMusic and in its API. It reports a blind-test preference across 47,885 pairs and sets lossless download limits by plan. Tracks built on other artists' songs cannot be downloaded.

ElevenLabs forward-deployed engineers describe how to ship voice agents that resolve issues rather than deflect them. They split agents into deterministic software and a core orchestrator, then cover versioning, A/B testing, governance and release cycles.

ElevenLabs outlines a voice agent evaluation framework across voice quality, conversation quality, tool use, intelligence, compliance and reliability, with production targets and testing pitfalls.

A build walkthrough of a mobile app that identifies statues from photos, designs period-appropriate voices per character, and runs a multi-voice real-time agent over WebRTC, including a single prompt that one-shots it in a coding agent.

ElevenLabs opens image, video and template APIs, exposing models from Google, OpenAI, ByteDance and Creatify at app pricing. Templates built in Flows can be run programmatically, and the same models are reachable through its MCP server.

Compares statistical and watermark-based audio deepfake detection, their strengths and blind spots, and how to combine them in large-scale fraud detection stacks for voice systems.

A recap of a workshop on bolting voice onto an existing chatbot: why silence-based turn detection fails, a dual WebSocket architecture that streams LLM output into synthesis, and how much infrastructure to own.