{"company":{"name":"Deepgram","slug":"deepgram","website":"https://deepgram.com","category":"ai-ml"},"question":"What has Deepgram shipped recently?","answer":"In the last 30 days, Deepgram shipped 41 tracked updates. The most recent was \"Deepgram introduces Keyterm Prompting to boost transcription accuracy for industry terms\" on 2026-10-02.","window":{"days":30,"updateCount":41,"returned":41},"generatedAt":"2026-10-03T21:20:15.129Z","updates":[{"title":"Deepgram introduces Keyterm Prompting to boost transcription accuracy for industry terms","summary":"Deepgram launched Keyterm Prompting, a feature that instantly improves recognition accuracy for up to 100 important terms, product names, and jargon in transcription tasks. Available for Nova-3 and Flux models, it replaces the legacy Keywords feature with a simpler syntax and supports multi-word phrases via URL encoding. Benchmark examples show confidence scores rising from ~0.8 to ~0.98 for misrecognized terms like 'tretinoin' or 'escalation'.","date":"2026-10-02","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://developers.deepgram.com/docs/keyterm.md","publisher":"developers.deepgram.com"},{"title":"Deepgram adds real-time stream configuration via 'Configure' message for Nova-3 models","summary":"Deepgram introduced a 'Configure' message for its `/v1/listen` WebSocket streaming API, allowing dynamic adjustments to keyterms and formatting features (e.g., numerals) mid-stream without dropping audio or rebuilding state. The feature is available globally for Nova-3 models but not yet on regional endpoints (EU, Australia, India).","date":"2026-10-02","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://developers.deepgram.com/docs/configure.md","publisher":"developers.deepgram.com"},{"title":"Deepgram adds mid-stream keyterms updates for Nova-3 models","summary":"Deepgram introduced the ability to dynamically swap keyterms mid-stream for Nova-3 models during active /v1/listen sessions using a Configure message. The update replaces the entire keyterms list, enforces the 500-token limit, and works across all Nova-3 variants except regional endpoints. The feature is documented in the /v1/listen API reference.","date":"2026-10-02","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/10/2.md","publisher":"developers.deepgram.com"},{"title":"Deepgram Self-Hosted 261001: Flux TTS watermarking, Whisper removal, FIPS updates","summary":"Deepgram’s October 2026 Self-Hosted release (261001) adds mandatory Flux TTS audio watermarking, removes Whisper model support, and introduces FIPS 140-3 compliant images with updated metrics and certificate endpoints. It also improves Japanese and German number/date formatting in Aura-2 and Flux TTS.","date":"2026-10-01","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/10/1.md","publisher":"developers.deepgram.com"},{"title":"Deepgram Flux TTS gains inline pause and pronunciation controls","summary":"Deepgram’s Flux TTS now supports inline pause insertion (500–3000 ms) and pronunciation overrides via IPA in the same request as speed adjustments. Pause and pronunciation cannot be combined with speed changes outside defaults, and invalid combinations are rejected with clear errors. The pronunciation feature is in Early Access and may vary between generations.","date":"2026-09-30","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/30.md","publisher":"developers.deepgram.com"},{"title":"Voice Agents Built for the Conversation: Building with Deepgram’s Flux TTS on Pipecat","summary":"Deepgram and Daily announced a live technical webinar on September 29, 2026, showcasing their combined stack for building real-time voice agents. The session will demonstrate Flux TTS and Pipecat’s integration, highlighting features like persistent tone across turns, interruption handling, and low-latency multi-turn conversations.","date":"2026-09-29","dateIsEstimated":false,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://deepgram.com/webinars/voice-agents-built-for-the-conversation","publisher":"deepgram.com"},{"title":"Deepgram adds Numerals feature for converting written numbers to numerical format in transcripts","summary":"Deepgram introduced a Numerals feature that converts written numbers (e.g., 'nine hundred') into numerical format (e.g., '900') in transcripts. The feature supports multiple languages, including English variants, Spanish, French, German, and others, with toggling available during real-time and Flux STT streams. It can be enabled via API query parameters or Configure messages.","date":"2026-09-28","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/docs/numerals.md","publisher":"developers.deepgram.com"},{"title":"Deepgram Flux gains mid-stream dynamic configuration for voice agents","summary":"Deepgram introduced a mid-stream `Configure` control message for its Flux speech recognition model, allowing real-time adjustments to keyterms, language hints, turn detection thresholds, and numerals without disconnecting. This enables voice agents to dynamically adapt ASR behavior to evolving conversation phases, reducing engineering complexity and improving accuracy for task-critical phrases.","date":"2026-09-28","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://developers.deepgram.com/docs/flux/configure.md","publisher":"developers.deepgram.com"},{"title":"Deepgram Flux STT gains mid-stream numeral toggling for 8 languages","summary":"Deepgram added the ability to toggle numerals (digits vs. words) mid-stream in Flux STT without reconnecting, enabling dynamic formatting for PINs, phone numbers, or order IDs. The feature works across eight languages and is controlled via a Configure message with a boolean `numerals` parameter, replacing the previous connection-time-only behavior.","date":"2026-09-25","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/25.md","publisher":"developers.deepgram.com"},{"title":"Deepgram Nova-3 models improved for Flemish, German (Switzerland), Lithuanian, and Portuguese","summary":"Deepgram released improved Nova-3 monolingual models for Flemish, German (Switzerland), Lithuanian, and Portuguese, enhancing transcription quality for batch and streaming workloads. No API changes are required; updates are live for all users, with self-hosted customers needing to request updated models.","date":"2026-09-24","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://developers.deepgram.com/changelog/2026/9/24.md","publisher":"developers.deepgram.com"},{"title":"Flowtica ships AI note-taking pen to turn transcripts into next steps","summary":"Flowtica’s AI note-taking pen, Flowtica Scribe, has shipped over 10,000 units and relies on Deepgram’s Speech-to-Text API for transcription. Deepgram’s batch transcription handles multi-speaker, multi-language audio in real-world conditions, enabling the pen’s action-item generation and subscription revenue.","date":"2026-09-23","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/customers/flowtica","publisher":"deepgram.com"},{"title":"Synthflow AI scales global voice agents with Deepgram’s real-time Speech AI","summary":"Deepgram’s Nova-3 and Flux models power Synthflow AI’s no-code voice agent platform, enabling real-time, multilingual interactions with low latency and code-switching support. Synthflow self-hosts Deepgram in the EU for data residency and performance, while leveraging a dedicated engineering partnership for optimization.","date":"2026-09-23","dateIsEstimated":true,"signalType":"case_study","signalTypeLabel":"Case Study","sourceUrl":"https://deepgram.com/customers/synthflow-ai","publisher":"deepgram.com"},{"title":"Prosper AI verifies patients and benefits accurately with Deepgram Nova-3","summary":"Deepgram’s Nova-3 speech-to-text model is now the transcription layer for Prosper AI’s healthcare voice agents, handling real-time streaming and batch transcription for benefits verification and billing workflows. Prosper AI reports 98.5% successful identity verifications and 99%+ structured data accuracy, enabling first-call resolution for payer and patient calls.","date":"2026-09-23","dateIsEstimated":true,"signalType":"case_study","signalTypeLabel":"Case Study","sourceUrl":"https://deepgram.com/customers/prosper-ai","publisher":"deepgram.com"},{"title":"Build the future ofVoice AI together","summary":"Deepgram unveiled a structured partner program with tracks for technology, channel, and consultancy partners, highlighting integrations with AWS (Generative AI Competency Partner) and Five9 (Nova-2 ASR in IVA Studio 7). Success stories showcase real-time STT and TTS deployments across financial, sales, and video platforms.","date":"2026-09-22","dateIsEstimated":true,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://deepgram.com/partners","publisher":"deepgram.com"},{"title":"The future is talking. Turn it up.","summary":"Deepgram announced Speak, a one-day voice AI conference in San Francisco on October 29 with 500 seats, featuring CEO Scott Stephenson and customers in production. The event targets AI decision-makers, developers, and press, with sponsors including NiCE, AWS, and Vapi.","date":"2026-09-22","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/speak","publisher":"deepgram.com"},{"title":"Deepgram Nova-3 models improved for nine languages","summary":"Deepgram released enhanced Nova-3 monolingual models for Danish, Estonian, Flemish, Italian, Lithuanian, Macedonian, Polish, Urdu, and Vietnamese, improving transcription accuracy for batch and streaming workloads. No API changes are required; updates are live for all users, with self-hosted customers needing to request updated models.","date":"2026-09-22","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/22.md","publisher":"developers.deepgram.com"},{"title":"Three acknowledgements you cannot wait for","summary":"Deepgram’s Voice Agent API (WebSocket) sends acknowledgments like ListenUpdated and ThinkUpdated after model switches, but these may not always arrive. The article describes a client-side pattern using a three-second timer and fallback reconnect to handle missing acks, ensuring seamless mid-call model changes (e.g., Flux STT to Nova-3).","date":"2026-09-21","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepgram.com/learn/voice-agent-updatelisten-no-ack-fallback","publisher":"deepgram.com"},{"title":"Deepgram launches Browser Agent SDK with four composable packages for web voice agents","summary":"Deepgram introduced a new Browser Agent SDK consisting of four composable JavaScript packages—Widget, React UI Components, React Hooks, and JavaScript SDK—designed to connect web apps to its Voice Agent API. The SDK supports any framework, offers six pre-built layouts, and includes features like reconnection logic, playback-aware mode tracking, and client-side function calling for React apps.","date":"2026-09-18","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/docs/browser-agent-overview.md","publisher":"developers.deepgram.com"},{"title":"Deepgram corrects Browser Agent SDK VAD and token-minting documentation","summary":"Deepgram’s changelog clarifies that the Browser Agent SDK never included client-side voice activity detection (VAD) despite documentation suggesting otherwise, and fixes a token-minting example where the `ttl` field should be `ttl_seconds`. The corrections affect microphone streaming, barge-in behavior, and token expiration logic.","date":"2026-09-18","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/18.md","publisher":"developers.deepgram.com"},{"title":"Introducing Nova-3 Pharma: The First STT Model Purpose-Built for the Pharmaceutical Industry","summary":"Deepgram introduced Nova-3 Pharma, the first speech-to-text model purpose-built for pharmaceutical applications, excelling in drug name recognition and overall transcription accuracy. The model outperforms competitors in benchmarks for both batch and streaming transcription across eight English variants.","date":"2026-09-17","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/learn/introducing-nova-3-pharma","publisher":"deepgram.com"},{"title":"Deepgram launches nova-3-pharma model for pharmaceutical speech-to-text","summary":"Deepgram introduced `nova-3-pharma`, a specialized Nova-3 model optimized for pharmaceutical vocabulary and drug-name recognition in English. The model supports both batch and streaming transcription and is available via API for hosted customers, with self-hosted access by request.","date":"2026-09-17","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/17.md","publisher":"developers.deepgram.com"},{"title":"Deepgram updates self-hosted deployment docs for NVIDIA drivers and container platforms","summary":"Deepgram revised its self-hosted deployment documentation to mandate NVIDIA driver version 580 or higher with open kernel modules for GPU support, replacing the previous 570.172.08 requirement. The update also formalizes Docker as the recommended containerization platform while acknowledging Podman as an alternative for RHEL-based systems.","date":"2026-09-16","dateIsEstimated":true,"signalType":"technical","signalTypeLabel":"Technical","sourceUrl":"https://developers.deepgram.com/docs/drivers-and-containerization-platforms.md","publisher":"developers.deepgram.com"},{"title":"Deepgram's India Endpoint Is Now Generally Available","summary":"Deepgram’s India endpoint is now generally available, enabling onshore speech-to-text, text-to-speech, and Voice Agent workloads with storage and inference in AWS’s ap-south-2 (Hyderabad) region. The fully managed service matches API parity with global endpoints and targets regulated sectors like banking, payments, and insurance.","date":"2026-09-15","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/learn/deepgram-india-endpoint-now-generally-available","publisher":"deepgram.com"},{"title":"Deepgram launches India endpoint for Speech-to-Text, TTS, Voice Agent, and Text Intelligence APIs","summary":"Deepgram announced general availability of its India endpoint (api.in.deepgram.com) supporting Speech-to-Text, Text-to-Speech, Voice Agent, and Text Intelligence APIs. Customers can now process data within India by replacing api.deepgram.com with api.in.deepgram.com in requests, with existing API keys and tokens remaining compatible.","date":"2026-09-15","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/15.md","publisher":"developers.deepgram.com"},{"title":"TTS with emotion: emphasis and control in live conversations","summary":"Deepgram’s Flux TTS now supports three ways to control emotional delivery in live conversations: inline markup from the LLM, a connection-level expressivity parameter (-2 to 2), or persistent state across turns. Non-default settings increase word-level risk but enable turn-by-turn expressivity. The expressivity parameter is beta and validated only at default for production use.","date":"2026-09-14","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/learn/tts-emotion-control-live-voice-agents","publisher":"deepgram.com"},{"title":"Deepgram SDKs add Voice Agent ForceEndTurn and Flux TTS expressivity/speed controls","summary":"Deepgram released updated Python, JavaScript, and Java SDKs (v7.9.0, v5.11.0, v0.10.0) adding Voice Agent ForceEndTurn support and Flux TTS expressivity controls (-2 to 2) with expanded speed range (0.5–1.5 in 0.05 increments). Java SDK v0.10.0 includes breaking changes removing deprecated constants.","date":"2026-09-14","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/14.md","publisher":"developers.deepgram.com"},{"title":"Build A Voice Agent With Pipecat And Deepgram Flux STT And Flux TTS","summary":"Deepgram introduced Flux TTS, a conversation-native text-to-speech model, alongside Flux STT, both now generally available. The models reduce pipeline complexity by combining transcription with turn detection, eliminating the need for separate VAD and turn analyzers. Flux TTS is free to use through September 12, 2026.","date":"2026-09-12","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/learn/build-voice-agent-pipecat-deepgram-flux-stt-tts","publisher":"deepgram.com"},{"title":"Deepgram CLI adds named profiles and enhanced authentication methods","summary":"Deepgram updated its CLI documentation to introduce named profiles for managing multiple API keys via the system keyring, alongside existing interactive login, direct API key flags, and environment variable methods. The CLI now supports profile-specific credentials and configuration, improving multi-environment workflows.","date":"2026-09-12","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/developer-tools/cli/authentication.md","publisher":"developers.deepgram.com"},{"title":"Deepgram CLI gains plugin system for extensibility","summary":"Deepgram’s `dg` CLI now supports a plugin system allowing users to install, list, update, and remove third-party commands via `dg plugin` commands. Plugins are Python packages exposing a `BaseCommand` class, registered via `deepctl.plugins` entry points in `pyproject.toml`.","date":"2026-09-12","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/developer-tools/cli/plugins.md","publisher":"developers.deepgram.com"},{"title":"Deepgram introduces temporary token-based authentication for APIs","summary":"Deepgram launched temporary tokens with a 30-second TTL for secure, client-side API access, replacing traditional API keys. Tokens are generated via a backend service using the `/auth/grant` endpoint and passed to clients for immediate use in real-time applications like WebSockets and REST calls.","date":"2026-09-12","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/guides/fundamentals/token-based-authentication.md","publisher":"developers.deepgram.com"},{"title":"Deepgram enables custom parameters in all SDKs for unsupported API features","summary":"Deepgram’s JavaScript, Python, .NET, Go, and Java SDKs now support passing arbitrary key/value pairs as custom parameters in API requests, even for features not officially defined in each SDK. This allows developers to leverage undocumented or emerging API capabilities immediately.","date":"2026-09-12","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/guides/fundamentals/using-custom-parameters-sdks.md","publisher":"developers.deepgram.com"},{"title":"Voice agent function calling: real-time data lookup during a call","summary":"Deepgram introduced function calling for voice agents to perform real-time data lookups during calls, addressing latency and failure modes. The update requires handling parameters from speech, managing dead air with filler functions, and ensuring conversational failure responses. The system supports client-side and server-side execution, with a focus on minimizing caller wait times.","date":"2026-09-10","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepgram.com/learn/voice-agent-function-calling-real-time-data-lookup","publisher":"deepgram.com"},{"title":"What is Speaker Diarization?","summary":"Deepgram defines speaker diarization as the process of identifying and separating individual speakers in audio streams, enabling structured conversation transcripts with speaker attribution. The post details how the technology works—through detection, segmentation, representation, and attribution—and highlights its business impact across contact centers, healthcare, legal, sales, media, education, and recruiting. Deepgram emphasizes its language-agnostic, real-time capabilities and unlimited speaker support.","date":"2026-09-10","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://deepgram.com/learn/what-is-speaker-diarization","publisher":"deepgram.com"},{"title":"The Rise of PER: A New Standard for Measuring TTS Accuracy","summary":"Deepgram proposes Pronunciation Error Rate (PER) as a new TTS evaluation metric to address gaps in Mean Opinion Score (MOS) and Word Error Rate (WER), which obscure pronunciation-specific failures. PER isolates phoneme-level errors, enabling targeted remediation for critical terms like brand names, alphanumeric IDs, and medical/financial terminology. The article details PER’s calculation, implementation approaches, and integration with existing metrics like WER and MOS.","date":"2026-09-10","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://deepgram.com/learn/per-new-standard-measuring-tts-accuracy","publisher":"deepgram.com"},{"title":"Voice agent testing: how to evaluate before you take real calls","summary":"Deepgram’s guide details a nine-scenario testing suite for voice agents, emphasizing whole-call scoring over transcript metrics like WER. It warns that synthetic TTS clips miss real-caller behaviors (overlaps, disfluencies) and recommends mining live call recordings to build robust test cases. Pass criteria focus on task completion, escalation, and false statements, with local thresholds set by business value.","date":"2026-09-10","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://deepgram.com/learn/voice-agent-testing-evaluate-before-real-calls","publisher":"deepgram.com"},{"title":"Speech prosody across languages: what transfers and what breaks","summary":"Deepgram expanded language support across its transcription (Nova-3), conversational recognition (Flux Multilingual), and synthesis (Aura-2) models. Nova-3 now transcribes Mandarin, Cantonese, and Thai, while Flux Multilingual covers 10 non-tonal languages (e.g., English, Spanish, Japanese). Aura-2 handles multilingual synthesis, with Flux TTS available only for English. Tone languages remain limited to transcription due to pitch's dual role in lexical meaning and turn detection.","date":"2026-09-10","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://deepgram.com/learn/speech-prosody-across-languages","publisher":"deepgram.com"},{"title":"Voice AI latency: why pauses read as hesitation","summary":"Deepgram’s guide reveals how voice AI latency—measured as the silence between user speech and agent response—directly shapes caller perception, with 61% of surveyed developers reporting flow issues. Human turn transitions average under 250ms, but conventional pipelines exceed 1,000ms, creating hesitation, uncertainty, or the impression of inattention. The post argues for deployment-specific latency budgets (e.g., 260ms for phone IVR, 1s+ for live captioning) and emphasizes tracking p90 silence rather than median performance.","date":"2026-09-10","dateIsEstimated":true,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://deepgram.com/learn/voice-ai-latency-why-pauses-read-as-hesitation","publisher":"deepgram.com"},{"title":"Voice AI telephony integration: carriers, SLAs and global rollout","summary":"Deepgram outlines two voice AI telephony integration paths: media streaming over WebSocket (via CPaaS like Twilio) and SBC/SIP hand-off for existing PBX/CCaaS systems. The article details failover, regional redundancy, and global deployment requirements, including country-specific number provisioning and data residency rules. Deepgram supports regional endpoints in North America, Europe, and Australia with updated concurrency limits as of February 2026.","date":"2026-09-10","dateIsEstimated":true,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://deepgram.com/learn/voice-ai-telephony-integration-carriers-slas-global-rollout","publisher":"deepgram.com"},{"title":"Deepgram releases @deepgram/react 0.2.0 with runtime controls and safer lifecycle handling","summary":"Deepgram shipped `@deepgram/react` 0.2.0, adding runtime Voice Agent controls (sendAgentMessage, updateListen/Think/Speak/Prompt) and safer React session lifecycle handling. The update includes typed callbacks for protocol notifications, a new 'thinking' mode, and improved cleanup for canceled or failed resources.","date":"2026-09-10","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/10.md","publisher":"developers.deepgram.com"},{"title":"Q&A with Deepgram’s New Chief Product Officer, Ed Anuff","summary":"Deepgram named Ed Anuff as its new Chief Product Officer, bringing over 30 years of experience in product leadership, including roles at IBM watsonx.data and Apigee. Anuff highlighted voice AI as the next major interaction layer for AI, positioning Deepgram to expand beyond speech-to-text into full voice agent systems.","date":"2026-09-08","dateIsEstimated":true,"signalType":"executive_change","signalTypeLabel":"Executive","sourceUrl":"https://deepgram.com/learn/deepgram-chief-product-officer-ed-anuff","publisher":"deepgram.com"},{"title":"Deepgram adds defer_until_eot and FunctionCallCancelled for voice agents","summary":"Deepgram introduced two new features to control function calls in voice agents: `defer_until_eot` to hold irreversible actions until user turn confirmation and `FunctionCallCancelled` to notify clients when a function call is discarded due to user barge-in. These changes apply across all listen providers, not just Flux.","date":"2026-09-08","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://developers.deepgram.com/changelog/2026/9/8.md","publisher":"developers.deepgram.com"}],"attribution":{"source":"Spyingbee","url":"https://spyingbee.com/updates/deepgram","citation":"Spyingbee, \"Deepgram updates\", https://spyingbee.com/updates/deepgram (retrieved 2026-10-03)"}}