AI Voice and Speech: Release Notes

ElevenAgents

• Structured procedures compile on save and publish: Update agent and Create agent draft now compile and validate structured procedures automatically. Integrations no longer need to call the compile endpoint or manually update the generated workflow. Invalid procedures return per-procedure errors as 400 procedure_validation_failed. Agents without structured procedures are unaffected.
• Compile procedures endpoint is legacy: Compile procedures is marked legacy and should not be used by new integrations. It remains available for existing callers.
• Merge proposals: Merge proposals are available in alpha. Teams can open a proposal for an agent branch, review configuration changes and test results, leave comments or approval decisions, and merge approved changes into a target branch. See Merge proposals.

SDK releases

JavaScript SDK …

Fish Audio S2.1 Pro

Production text-to-speech model that improves on S2-Pro in quality, latency, and throughput. Supports 83 languages on one model, with the same [bracket] natural-language control and multi-speaker dialogue as S2.

Use model ID s2.1-pro for production workloads that need TTFA and DPA guarantees. s2.1-pro-free is the same model at no cost through November 30, 2026 for testing, prototyping, development, and smaller businesses, under fair-use limits and without those guarantees.

Read more about S2.1 Pro | Models overview

Eleven v4 and Eleven v4 Turbo

Eleven v4 and Eleven v4 Turbo are now available. Eleven v4 provides higher-quality, more expressive speech and improved voice cloning across more than 90 languages. Eleven v4 Turbo provides the same model family for real-time applications, with median inference latency of approximately 100 ms.

Use eleven_v4 through the Text to Dialogue API for content creation and long-form audio. Use eleven_v4_turbo through the Text to Dialogue WebSocket for agents and interactive applications.

Speech to Text

• Transcript editing: Batch and realtime transcription now accept a natural-language edit instruction of up to 2,000 characters. Batch responses return the result in edited_transcript, while realtime sessions emit an edited_transcript event. See the batch and realtime guides for incompatibilities, response handling and pricing.

ElevenAgents

• Agent models: Added gpt-6-sol, gpt-6-luna, glm-52, deepseek-v41-flash, claude-opus-5 and claude-opus-5-5 to the supported agent LLM options.
• Agent test usage: Test invocation summaries can include total credits and USD price. Individual test runs can include credits and a charging breakdown.

Flows …

ElevenAgents

• Parallel tool calls: Agent prompt configuration now includes enable_parallel_tool_calls (boolean, default true). When enabled, supported models can execute multiple tools in one turn.
• Agent models: Added gpt-6-astra to the agent LLM options.
• Alerting: Agent alerting now supports Slack channels alongside PagerDuty and webhook notifiers.
• Phone number search: Added a cursor-paginated phone number endpoint with filters for provider, outbound support, assigned agent and branch, label and phone number.

Image and video

• GPT Image 2.5 models: Image generation now supports gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both models accept up to 10 reference images, quality levels through max, 14 fixed aspect ratios plus auto, and 1K, 2K or 4K output.

Speech Engine

• Cascade timeout: Speech Engine create and update requests now accept cascade_timeout_seconds (number, 2–15, default 4) to control how long ElevenLabs waits for an upstream speech engine before retrying.

Text to Speech

• Multi-context turn boundaries: The is_final_audio_for_turn marker is now emitted after every buffered byte for that turn, including MP3 and Opus output. Clients can use the marker as an exact turn boundary without switching to PCM.

SDK releases

JavaScript SDK …

Flow now runs on Canto, the first speech model we've built ourselves. It's now rolled out for everyone, with nothing to switch on.

Most speech models are trained and tested on clean audio recorded in quiet rooms with good microphones. That isn't where people use Flow. You're on a train, in an open office, walking between meetings, or talking quietly so you don't disturb anyone. Canto was built and measured against those conditions instead.

On our evaluation of real-world English Flow recordings, 9.8 hours from more than 2,300 speakers, Canto had the lowest word error rate of every model we tested, including models from Google, OpenAI, AssemblyAI and Deepgram.

On a separate set built to stress-test the hardest audio, with background conversation, music, traffic, wind and whispered speech, Canto was the most accurate of the real-time models we tested.

We're already training its successor at more than ten times the scale, with better recognition in difficult conditions and broader language coverage ahead.

Read the full research post here.

‍

September 15, 2026

Notetaker is now available on Windows, alongside Mac. Here’s what it does and what we’ve added since the Mac launch.

Capture the details, then work from them

Notetaker detects meetings across the platforms you already use and offers to take notes. It uses your personal dictionary to help capture names, acronyms, and technical terms correctly so you can:

• Catch up during a call: Ask “What did I miss?” for a recap of the discussion so far.
• Find answers across meetings: Ask “What follow-ups do I owe people from my calls this week?” or “Based on the calls I’ve led this month, how could I ask better questions?” Answers include citations back to the meetings.
• Draft from your notes: Turn conversations into follow-up emails, project updates, or briefs directly in Notetaker.

Find it: Open the Notetaker tab in Wispr Flow.

Connect your meetings to your other work

Notetaker just added one-click setup for Gemini, alongside connections to Claude and ChatGPT. You can also connect to any other MCP-compatible AI tool.

Once connected, your AI tool can grab your meeting notes and transcripts along with any other files and tools you’ve connected there. You’ll be able to dictate complex prompts:

• For sales: “Compare the objections from my discovery calls this quarter with our objection-handling doc. Where are we not answering well?”…

ElevenAgents

• Call queueing and hold audio: Added call queueing for agents at their concurrency limit. Queued callers hear hold audio and receive queue_status events until they are admitted or time out. New API methods upload or delete custom hold audio.
• Test version metadata: Agent test results now identify the agent version they ran against and whether they included draft changes. Invocation summaries also indicate when partial resubmissions ran against different versions.
• Model and account metadata: Agent LLM configuration now accepts gemini-3.8-flash. WhatsApp account responses identify whether the account uses the Cloud API or coexistence signup flow.

Music

• Music v2.5 API support: Music endpoints and SDKs now accept music_v2_5. Composition plan generation chunks support up to 6,132 characters, with up to 30 lines of 200 characters each.

SDK releases

JavaScript SDK

• v2.68.0 - Added methods to upload and delete agent hold audio. Regenerated types for call queueing, procedure references, draft branch creation, RAG query limits, test version metadata, gemini-3.8-flash and music_v2_5.

Python SDK …

OpenAI-compatible speech endpoint

Realtime TTS now serves POST /v1/audio/speech in OpenAI's format, so apps built on OpenAI's text-to-speech can switch to Inworld by pointing the SDK at Inworld and choosing an Inworld model and voice — see OpenAI Compatibility.

• Works with the OpenAI SDKs: The official Python and Node.js SDKs work unchanged, including streaming responses.
• Formats: mp3, opus, flac, wav, and pcm, plus server-sent events with stream_format: "sse".
• One base URL: https://api.inworld.ai/v1 also serves the Inworld Router, so one client configuration covers LLM and TTS.

ElevenAgents

• Workspace-wide conversation tickets: Added List workspace tickets to retrieve conversation triage tickets across accessible agents. Filter by status or assignee and use cursor pagination to retrieve up to 100 tickets per page.
• Dynamic variable conversation filters: List conversations and Text search conversation messages now accept repeatable dynamic_variable_params filters. Use name:op:value with eq, gt, gte, lt or lte; comparison operators require numeric values.
• Twilio answering machine detection: Outbound telephony and batch call configurations now support optional twilio_machine_detection. Choose enable for an early human-or-machine verdict or detect_message_end to wait for a voicemail greeting to finish. Results arrive through the new answering_machine_detection webhook event.
• Data collection value constraints: Data collection properties can now use allowed_values to name a dynamic variable containing the permitted values. The previous allowed_values_dynamic_variable field is deprecated.
• Agent metadata: List agent branches responses now indicate when a draft was created and whether it predates the branch tip. Agent and Speech Engine summaries now require voice_id. …

You can now bring your Otter.ai meeting transcripts straight into Wispr Flow. Just drop in your Otter export and your meetings appear alongside everything else, with speaker names preserved instead of collapsed into a single anonymous voice. Imports run in the background, so you can keep working while your history fills in.

Find it: Settings > Notetaker > Import, or the import picker where you'll see the Otter tile alongside Granola.

[Help Center: Import your Otter transcripts into Wispr Flow Notetaker]

For Teams & Enterprise

• Assign teams at invite time: Admins can now pick which team a new member joins right from the invite form, for single invites or bulk batches, so new hires land in the right place from day one.
• Org-wide and department-wide dictionary words: You can now choose to share dictionary entries to either your whole organization or your department.

[Help Center: Manage Users and Billing]

Quality-of-Life Improvements

• Updates wait until you're done: Flow no longer restarts itself mid-task. It detects when you're actively working and holds pending updates until you step away.

‍

v2.3.2