{"company":{"name":"AssemblyAI","slug":"assemblyai","website":"https://www.assemblyai.com","category":"ai-ml"},"question":"What has AssemblyAI shipped recently?","answer":"In the last 30 days, AssemblyAI shipped 59 tracked updates. The most recent was \"Business use cases for Generative AI\" on 2026-11-19.","window":{"days":30,"updateCount":59,"returned":50},"generatedAt":"2026-11-22T11:32:50.931Z","updates":[{"title":"Business use cases for Generative AI","summary":"AssemblyAI published a guide to business generative AI use cases across healthcare, customer service, financial services, content, education and legal. It highlights its LLM Gateway with 55+ models and automatic fallback, the AssemblyAI-hosted Qwen3.5 4B Fast model at $0.10/$0.50 per million tokens with 612 ms average voice rewrite latency, and Medical Mode at $0.15/hr with a 3.2% Missed Entity Rate. The post also cites customer results from Siro and Calabrio and states HIPAA BAA availability.","date":"2026-11-19","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://www.assemblyai.com/blog/business-use-cases-generative-ai","publisher":"assemblyai.com"},{"title":"How to measure transcription accuracy and evaluate speech recognition models","summary":"AssemblyAI published a guide on evaluating speech recognition models beyond word error rate, covering semantic WER, missed entity rate, CER, cpWER, and a benchmarking framework with CI-based WER regression testing. The post cites its own models, including AssemblyAI-STT-3.7 priced at $0.21/hr for pre-recorded audio, and its Truth File Corrector tool. It also states that a Universal-3.5 Pro average of 4.35% WER across four dataset groups.","date":"2026-11-19","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://www.assemblyai.com/blog/how-to-evaluate-speech-recognition-models","publisher":"assemblyai.com"},{"title":"Speech Understanding tasks explained: Speaker ID, custom formatting, and translation","summary":"AssemblyAI's blog describes its Speech Understanding tasks, which turn raw transcripts into structured data. Advanced speaker identification offers role-based labeling (for example agent versus customer) and name-based labeling built on diarization. Custom formatting applies rules for dates, phone numbers, emails, URLs and currency. Integrated translation returns translated_text for up to 86 target languages in the same request as transcription, supported on AssemblyAI-STT-3.7 and Universal-2. The article also notes real-time translation over the streaming WebSocket using AssemblyAI-STT-3.7-Realtime and LLM Gateway, though the excerpt is truncated.","date":"2026-11-19","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/speech-understanding-tasks-explained-speaker-id-custom-formatting-translation","publisher":"assemblyai.com"},{"title":"AssemblyAI-STT-3.7: for pre-recorded audio, with 32 languages and powerful speaker diarization","summary":"AssemblyAI released AssemblyAI-STT-3.7 for pre-recorded audio, replacing the Universal model family, with language coverage expanded from 18 to 32 languages, native code switching, and a new speaker diarization model. The company reports cpWER down 20%, DER down 36% and speaker-count errors down 63% versus Universal-3.5 Pro at launch, with 3–4× faster processing, all at the unchanged price of $0.21/hr. The diarization model is live across all async speech models, and a new pair of request parameters lets customers control verbatim and disfluency output.","date":"2026-11-18","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://www.assemblyai.com/blog/assemblyai-stt-3-7-research","publisher":"assemblyai.com"},{"title":"Introducing AssemblyAI-STT-3.7: next generation transcription models for live conversations, voice agents, and recordings","summary":"AssemblyAI released AssemblyAI-STT-3.7, a family of three transcription models: AssemblyAI-STT-3.7-Realtime for live conversations and meetings, AssemblyAI-STT-3.7-Realtime-Agent for voice agents, and AssemblyAI-STT-3.7 for recorded audio. The company claims lower live diarization error than ElevenLabs Scribe v3 Live and Meta Muse Voice Transcribe, a 3.23% short-form English word error rate, and a final transcript returned 86 ms after speech ends for the agent model. The Realtime model is priced at $0.30 per hour, one-third less than the agent model, and merges previous streaming models into one 32-language stream.","date":"2026-11-18","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://www.assemblyai.com/blog/introducing-assemblyai-stt-3-7","publisher":"assemblyai.com"},{"title":"The most innovative companies with between 51 and 200 employees for 2025","summary":"Fast Company's 2025 list of the most innovative companies with 51 to 200 employees includes AssemblyAI for its transcription accuracy. The profile cites Universal-1 (spring 2024), its 22% accuracy edge over Azure, AWS, and Google, the follow-up Universal-2 model, and new 2024 pricing tiers from Nano at $0.12/hour to Best at $0.38/hour. The company reports doubled year-over-year revenue, 200,000 developers, and more than 5,000 paying organizations.","date":"2026-11-18","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.fastcompany.com/91270150/small-mighty-51-200-employees-most-innovative-companies-2025","publisher":"fastcompany.com"},{"title":"How to apply LLMs to multi-speaker audio recordings","summary":"AssemblyAI published a tutorial on pointing LLMs at multi-speaker recordings by first running speaker diarization so each utterance carries a speaker label. The post covers two paths: a one-shot query through LLM Gateway that pulls a transcript by ID, and a retrieval pipeline built with the Haystack framework and AssemblyAI's Transcriber component, which wires speaker labels into chunking so the speaker letter reaches the model. It also describes diarization parameters (speaker count ranges, speaker identification by name or role) and notes that Speaker Identification is a newer capability.","date":"2026-11-18","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/apply-llms-multi-speaker-audio","publisher":"assemblyai.com"},{"title":"How to build real-time translated captions for meetings (every speaker, every language)","summary":"AssemblyAI published a tutorial on building live, speaker-labeled meeting captions translated into multiple languages from one audio stream, using its AssemblyAI-STT-3.7-Realtime model with LLM Gateway for translation. The guide walks through a Python pipeline (companion repo) that fans captions out to viewers by language and exports per-language WebVTT files at meeting end. It states the realtime model costs $0.30/hr across 32 languages, and that translations typically land within about 100 ms of the final turn with a small model. It compares the approach against built-in Teams and Zoom translated captions, which are limited to the meeting client.","date":"2026-11-18","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://www.assemblyai.com/blog/real-time-translated-captions-meetings","publisher":"assemblyai.com"},{"title":"How to use the Kimi K3 API (pricing, context window, quickstart)","summary":"AssemblyAI published a guide on calling Moonshot AI's Kimi K3 through its LLM Gateway, using model ID kimi-k3 with an OpenAI-compatible endpoint. The Gateway lists $3 per 1M input tokens and $15 per 1M output tokens with global routing, a 1,000,000-token context window, and a 10% surcharge for US-region requests. The post also reports Voice Standard Bench results from an October 29, 2026 run, where Kimi K3 scored well on summaries and up to 15-minute transcripts but weakly on hour-long clean-up and slow latency, and it notes zero data retention on AWS Bedrock and Fireworks AI with no EU residency option.","date":"2026-11-17","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/kimi-k3-api","publisher":"assemblyai.com"},{"title":"AssemblyAI-STT-3.7-Realtime-Agent: speech-to-text built for voice agents","summary":"AssemblyAI released AssemblyAI-STT-3.7-Realtime-Agent, a streaming speech-to-text model built for voice agents. On a 12,460-call English voice-agent benchmark, the company reports the lowest word error rate and entity error rate among 26 streaming models tested, ahead of Soniox Realtime v6, ElevenLabs Scribe v3 Live and Deepgram Nova-3. The model returns the final transcript 86 ms after the caller stops speaking, half the 172 ms of its predecessor Universal-3.6 Pro Realtime. Compared with that predecessor, AssemblyAI reports lower word error rate, 54% lower background-speech leakage, and 25% to 38% lower short-response error rate. The model is trained with conversation and acoustic context and needs no speaker enrollment or separate voice-isolation step. These figures are AssemblyAI's own benchmark results.","date":"2026-11-17","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/assemblyai-stt-3-7-realtime-agent-research","publisher":"assemblyai.com"},{"title":"AssemblyAI updates dictation API with automatic fallback for recordings over two minutes","summary":"AssemblyAI’s dictation API now automatically falls back from the live endpoint (`/v1/transcribe/live`) to the finished-recording endpoint (`/v1/transcribe`) when audio exceeds 120 seconds, triggered by a 413 error with `error_code: \"audio_too_large\"`. The system preserves the full recording for transcription, with a 20-minute cap, and supports compressed audio formats (Ogg/Opus, FLAC) to reduce upload latency.","date":"2026-11-11","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/dictation/recordings-over-two-minutes","publisher":"assemblyai.com"},{"title":"AssemblyAI adds LLM-powered transcript rewriting to Dictation","summary":"AssemblyAI’s Dictation now supports real-time transcript rewriting via LLM instructions, returning both verbatim and cleaned-up versions in a single call. Users can customize rewrites with plain-English instructions or use the LLM Gateway for advanced control, including spoken punctuation handling and structured outputs like bulleted lists or clinical notes.","date":"2026-11-11","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/dictation/transcript-rewriting","publisher":"assemblyai.com"},{"title":"AssemblyAI adds code-switching to Universal-3.5 Pro for 18 languages","summary":"AssemblyAI introduced native code-switching detection in its Universal-3.5 Pro model, enabling real-time transcription of audio where speakers switch between languages mid-sentence across 18 supported languages. Users can enable this via the `language_detection` parameter or constrain the model to specific language sets using `language_codes`.","date":"2026-11-10","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/pre-recorded-audio/code-switching","publisher":"assemblyai.com"},{"title":"Medical Mode: now with fewer medical-term errors in pre-recorded and real-time transcription","summary":"AssemblyAI’s Medical Mode now cuts drug-name misses by 30% in English and 62% in Spanish for pre-recorded audio, while reducing overall medical-term errors across nine datasets. Real-time transcription also improves, with drug-name misses down 16% in English and 49% in Spanish. WER remains unchanged.","date":"2026-11-10","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/assemblyai-for-healthcare-research","publisher":"assemblyai.com"},{"title":"AssemblyAI for Healthcare: the most accurate voice AI for clinical products","summary":"AssemblyAI introduced an updated Medical Mode and a new Vet Mode (beta) to improve accuracy in clinical and veterinary speech-to-text. Medical Mode reduces missed medical terms by up to 27% in benchmarks, while Vet Mode targets veterinary-specific terminology errors. Both modes are available via API parameters and integrate with existing healthcare workflows.","date":"2026-11-10","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/assemblyai-for-healthcare","publisher":"assemblyai.com"},{"title":"Best real-time speech translation APIs in 2026","summary":"AssemblyAI now supports real-time translation within its Universal-3.6 Pro Realtime API by adding the llm_gateway connection parameter. Translations are returned as text on the same WebSocket alongside the source transcript, covering 32 languages with native code-switching. Users can choose their translation model and prompt, and pair the output with AssemblyAI-TTS-1 for speech in six languages.","date":"2026-11-10","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/best-real-time-speech-translation-apis","publisher":"assemblyai.com"},{"title":"How to make voice agents say numbers, emails, addresses and names correctly","summary":"AssemblyAI’s Text-to-Speech API (AssemblyAI-TTS-1) now normalizes common entities like phone numbers, prices, dates, emails, and addresses server-side without SSML or a pronunciation dictionary. A new guide explains how to format text for specific readings (e.g., dashes for digit-by-digit pronunciation) and integrate these rules into LLM prompts for voice agents.","date":"2026-11-10","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/voice-agent-pronunciation-numbers-emails-names","publisher":"assemblyai.com"},{"title":"How to build live translation into smart glasses and wearables","summary":"AssemblyAI published a technical guide explaining how to build live translation into smart glasses and wearables, focusing on a phone relay that streams audio to Universal-3.6 Pro Realtime with real-time translation enabled. The post highlights design decisions unique to wearables, such as handling partial vs. final translations, caption layout for small displays, and reconnection strategies for flaky connectivity.","date":"2026-11-10","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/live-translation-smart-glasses-wearables","publisher":"assemblyai.com"},{"title":"Real-time speech translation: how it works and how to build it","summary":"AssemblyAI introduced real-time speech translation for its Universal-3.6 Pro Realtime product, enabling developers to translate spoken language into another language or text in real time. The feature uses a cascaded pipeline (STT → MT/LLM → optional TTS) with latency budgets per stage, translating only final turns to avoid instability. A Python build demonstrates integration via the LLM Gateway, which is billed per token.","date":"2026-11-10","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/real-time-speech-translation","publisher":"assemblyai.com"},{"title":"Every feature, on one API request","summary":"AssemblyAI now enables any combination of 13 Speech Understanding features—including speaker diarization, identification, summarization, sentiment analysis, PII redaction, and translation—to run on the same transcription request via a single API call. Features are priced per hour of audio with incremental rates.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/features","publisher":"assemblyai.com"},{"title":"AssemblyAI adds speaker identification to transcription with name or role labels","summary":"AssemblyAI introduced a speaker identification feature that replaces generic labels like 'Speaker A' with real names or roles in transcripts without requiring voice enrollment. The feature relies on existing speaker diarization and can infer names from context or apply role-based labels like 'Interviewer' or 'Interviewee'.","date":"2026-11-08","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/speaker-identification","publisher":"assemblyai.com"},{"title":"AssemblyAI adds action items extraction to speech understanding API","summary":"AssemblyAI introduced a new action items extraction feature in beta for its speech understanding API, enabling users to generate timestamped action items from audio transcripts with optional effort levels (low or medium). The feature supports US and EU regions and can include decisions or adjust processing effort for higher precision.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/action-items","publisher":"assemblyai.com"},{"title":"AssemblyAI adds sentiment analysis to transcription API with speaker labels","summary":"AssemblyAI’s Sentiment Analysis model now detects positive, negative, or neutral sentiment for each spoken sentence in transcripts, returning confidence scores and timestamps. The feature is enabled via the `sentiment_analysis` parameter in the transcription request and supports US and EU regions. Speaker labels can be added by enabling `speaker_labels`, linking sentiment results to identified speakers.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/sentiment-analysis","publisher":"assemblyai.com"},{"title":"AssemblyAI adds Key Phrases feature to extract significant concepts from transcripts","summary":"AssemblyAI introduced a Key Phrases model that automatically identifies and extracts the most important words and phrases from transcripts, returning them in a separate `auto_highlights_result` object with relevancy ranks, occurrence counts, and timestamps. The feature is enabled via the `auto_highlights` parameter in the transcription API and is supported across multiple languages, models, and regions (US & EU).","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/key-phrases","publisher":"assemblyai.com"},{"title":"AssemblyAI adds Entity Detection to speech-to-text API","summary":"AssemblyAI introduced Entity Detection, a feature that automatically identifies and categorizes entities in transcribed audio, including people, organizations, addresses, phone numbers, medical data, and social security numbers. The feature is available in multiple languages and regions (US & EU) and can be enabled via the `entity_detection` parameter in their transcription API or SDKs.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/entity-detection","publisher":"assemblyai.com"},{"title":"AssemblyAI adds Topic Detection to transcripts via IAB Content Taxonomy","summary":"AssemblyAI introduced Topic Detection, a feature that identifies and categorizes topics in audio transcripts using the 698-topic IAB Content Taxonomy. The capability is enabled by setting `iab_categories` to true in transcription requests and is supported in both the API and SDKs across Python and JavaScript. Users receive timestamped topic labels with relevance scores and a summary of all detected topics.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/topic-detection","publisher":"assemblyai.com"},{"title":"AssemblyAI adds Custom Formatting to Speech Understanding for standardized transcript output","summary":"AssemblyAI introduced Custom Formatting within its Speech Understanding feature, enabling automatic standardization of dates, phone numbers, emails, currencies, and URLs in transcripts. Users can apply formatting rules (e.g., US/EU date styles, phone number patterns) either during transcription or as a post-processing step via a separate API call.","date":"2026-11-08","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/speech-understanding/custom-formatting","publisher":"assemblyai.com"},{"title":"Text-to-speech night inNew York","summary":"AssemblyAI announced an upcoming in-person event in New York focused on text-to-speech and voice agents, scheduled for December 2026. The event targets developers and companies building with voice AI, aiming to foster community engagement around AssemblyAI’s new TTS model.","date":"2026-11-08","dateIsEstimated":true,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://www.assemblyai.com/events/nyc-text-to-speech","publisher":"assemblyai.com"},{"title":"Text-to-Screech: Haunted TTS Launch Event","summary":"AssemblyAI unveiled its new text-to-speech model at a Halloween-themed launch event featuring life-size animatronics powered by real-time TTS, networking, and a costume contest. The event combined spooky entertainment with Voice AI demonstrations.","date":"2026-11-08","dateIsEstimated":true,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://luma.com/assemb-42oc","publisher":"luma.com"},{"title":"Hardware hackathon: turn a vintage toy into a voice agent","summary":"AssemblyAI is running an invite-only hackathon where participants will build a voice agent using its Voice Agent API and install it in a Furby toy. Attendees receive $100 in credits, a Furby keychain, and hands-on experience with real-time voice AI integration.","date":"2026-11-08","dateIsEstimated":true,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://luma.com/w9e4qgol","publisher":"luma.com"},{"title":"London Voice AI Meetup: Build your own voice app","summary":"AssemblyAI is hosting a London meetup to teach developers how to build a dictation app in 30 minutes using its API, which supports 18 languages and removes filler words. Attendees receive $100 in API credits and hands-on guidance from AssemblyAI engineers.","date":"2026-11-08","dateIsEstimated":true,"signalType":"event","signalTypeLabel":"Event","sourceUrl":"https://luma.com/k74g72a0","publisher":"luma.com"},{"title":"Understanding and comparing costs of text-to-speech models for voice agents","summary":"AssemblyAI published a cost comparison for TTS models used in voice agents, normalizing pricing to 600 characters per minute and highlighting hidden concurrency fees. It claims AssemblyAI-TTS-1 offers frontier naturalness, ~90ms TTFA, unlimited concurrency, and one-third the per-minute price of competitors like ElevenLabs and Cartesia.","date":"2026-11-06","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/tts-cost-comparison-voice-agents","publisher":"assemblyai.com"},{"title":"How to generate speech with Fish Audio: API, Python SDK, and self-hosted S2-Pro","summary":"AssemblyAI published a detailed guide comparing its AssemblyAI-TTS-1 model with Fish Audio’s offerings, emphasizing AssemblyAI-TTS-1’s lower latency (159ms vs. 321ms), higher accuracy in reading codes/addresses (92% vs. 86%), and simpler commercial terms ($15 per 1M characters with no prepaid tiers). The guide also highlights AssemblyAI-TTS-1’s unlimited concurrency and built-in voice cloning.","date":"2026-11-03","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/how-to-generate-speech-fish-audio","publisher":"assemblyai.com"},{"title":"How to choose a voice for your AI voice agent","summary":"AssemblyAI released a detailed guide on selecting the right voice for AI voice agents, emphasizing product-focused criteria like caller persona, brand alignment, and 8 kHz phone-line compatibility. It highlights AssemblyAI-TTS-1’s thousands of voices across six languages and provides a five-step testing process to ensure clarity in real-world call scenarios.","date":"2026-11-03","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/how-to-choose-a-voice-for-ai-voice-agent","publisher":"assemblyai.com"},{"title":"How to write text that makes TTS sound human: fillers, false starts and pacing","summary":"AssemblyAI published a guide on optimizing text input for its TTS API to sound more human-like, emphasizing sentence structure, fillers, false starts, and pacing over markup. The update applies to AssemblyAI-TTS-1 and includes system prompts for LLMs to generate replies written to be spoken, with no SSML or tags required.","date":"2026-11-03","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/tts-prompting-natural-speech","publisher":"assemblyai.com"},{"title":"TTS word timestamps: build read-along highlighting with word boundaries","summary":"AssemblyAI’s Text-to-Speech API now provides per-word timestamps (word boundaries) to enable precise read-along highlighting and live captions that stay synchronized across requests and interruptions. The feature is enabled by default on raw WebSocket connections and available via HTTP streaming, with SDK support for Python and JavaScript. Developers can anchor word timings to each request’s audio start time to avoid drift.","date":"2026-11-03","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/tts-word-timestamps","publisher":"assemblyai.com"},{"title":"AssemblyAI LLM Gateway introduces tiered rate limits with RPM and TPM tracking","summary":"AssemblyAI’s LLM Gateway now enforces tiered rate limits based on requests per minute (RPM) and tokens per minute (TPM), with limits tied to spend tiers and model-specific thresholds. New accounts start at Tier 0 (30 RPM, no TPM), while higher tiers (up to 3,000 RPM) unlock as spend grows. Limits are tracked over a rolling 60-second window, and token reservations are held for the duration of requests, even if unused.","date":"2026-11-02","dateIsEstimated":true,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/docs/llm-gateway/rate-limits","publisher":"assemblyai.com"},{"title":"One stream, two jobs: introducing SpeakerRevision","summary":"AssemblyAI introduced SpeakerRevision, a feature that revises speaker labels at the end of a live stream using full-context analysis, achieving async-grade accuracy (~30.31% cpWER) with only ~400ms added latency. The feature eliminates the need for separate async pipelines while maintaining real-time speaker labels during the stream.","date":"2026-10-30","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/introducing-speakerrevision","publisher":"assemblyai.com"},{"title":"Building a voice agent with a coding agent: why this approach beats a visual builder","summary":"AssemblyAI argues that using a coding agent (e.g., Claude Code) to build voice agents via its APIs is superior to visual builders. The post details a spec-first workflow, cascaded architecture (speech-to-text, LLM, TTS), and iterative development using AssemblyAI’s Universal-3.6 Pro Realtime, LLM Gateway, and Text-to-Speech APIs.","date":"2026-10-30","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://www.assemblyai.com/blog/building-a-voice-agent-with-a-coding-agent","publisher":"assemblyai.com"},{"title":"Sync HTTP for voice agents: where it helps, and why streaming stays the default","summary":"AssemblyAI emphasizes streaming as the default for voice agents but introduces three new Sync API use cases: second-pass transcription for low-confidence turns, audio recovery after WebSocket drops, and primary STT for low-volume agents with existing turn detection. Sync returns transcripts in ~134ms and costs $0.30/hr.","date":"2026-10-30","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/sync-http-voice-agents","publisher":"assemblyai.com"},{"title":"Claude Opus 5.5 vs GPT-6 Astra vs Gemini 4 Argon: which LLM wins on voice data?","summary":"AssemblyAI published an October 29, 2026 benchmark of three flagship LLMs on call transcripts, scoring summarization, extraction, correction and dictation alongside cost per audio hour and latency. Claude Opus 5.5 led summarization, GPT-6 Astra led long-transcript correction and Gemini 4 Argon led dictation and speed at the lowest cost. The post also finds that smaller models from the same families often come within a few points of their flagships.","date":"2026-10-29","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://www.assemblyai.com/blog/gemini-3-pro-vs-gpt-5-vs-claude-4-5","publisher":"assemblyai.com"},{"title":"Best Fish Audio alternatives for text-to-speech in 2026 (including ElevenLabs)","summary":"AssemblyAI published a competitive comparison positioning its AssemblyAI-TTS-1 model as a superior Fish Audio alternative for voice agents, emphasizing unlimited concurrency, faster first-audio latency (159ms vs 321ms), and 92% code accuracy vs Fish’s 86%. Pricing is $15 per 1M characters for both.","date":"2026-10-29","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/fish-audio-alternatives","publisher":"assemblyai.com"},{"title":"AssemblyAI updates TTS API with streaming and WebSocket support","summary":"AssemblyAI expanded its Text-to-Speech (TTS) API with a Streaming TTS feature that synthesizes audio in real-time and returns it via WebSocket for voice agents. The update includes new SDK versions (Python 1.8.2+, JavaScript 4.42.0+) and supports both HTTP and WebSocket endpoints, with data residency options for US and EU regions.","date":"2026-10-29","dateIsEstimated":true,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/docs/tts/getting-started/quickstart","publisher":"assemblyai.com"},{"title":"Build an Agora voice agent with AssemblyAI speech-to-text and text-to-speech","summary":"AssemblyAI announced a new integration with Agora’s RTC platform, enabling developers to build server-side voice agents that transcribe, process via LLM Gateway, and respond with text-to-speech—all using a single AssemblyAI API key. The tutorial demonstrates a Python bot that joins Agora channels, handles barge-in, and delivers real-time conversational responses with low latency.","date":"2026-10-29","dateIsEstimated":false,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://www.assemblyai.com/blog/build-agora-voice-agent-quickstart","publisher":"assemblyai.com"},{"title":"How to build a voice agent with AssemblyAI: streaming STT, LLM, and TTS","summary":"AssemblyAI now offers a unified voice agent stack with Universal-3.6 Pro Realtime for streaming speech-to-text, LLM Gateway for tool-calling LLMs, and Text-to-Speech, all accessible via a single API key. The stack includes built-in turn detection, interruption handling, and multilingual support (32 languages for STT, 6 for TTS), with accuracy benchmarks showing a 5.19% word error rate and 14.4% entity error rate versus competitors.","date":"2026-10-29","dateIsEstimated":false,"signalType":"product_launch","signalTypeLabel":"Launch","sourceUrl":"https://www.assemblyai.com/blog/how-to-build-a-voice-agent-with-assemblyai","publisher":"assemblyai.com"},{"title":"How to create an AI cold-calling agent with AssemblyAI","summary":"AssemblyAI published a guide to building an AI cold-calling agent using their Universal-3.6 Pro Realtime, LLM Gateway, and Text-to-Speech services. The agent qualifies prospects via structured scripts, writes CRM outcomes via tool calls, and includes a fail-closed compliance gate before dialing. Twilio handles telephony integration.","date":"2026-10-29","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/build-cold-calling-voice-agent-assemblyai","publisher":"assemblyai.com"},{"title":"Build a real-time voice agent in Python with tool calling","summary":"AssemblyAI introduced tool calling for real-time voice agents in Python, allowing developers to build agents that listen, call custom Python functions, and respond via text-to-speech using a single API key. The update leverages Universal-3.6 Pro Realtime for streaming speech-to-text, LLM Gateway for decision-making, and AssemblyAI’s Text-to-Speech API for high-quality, low-latency responses with barge-in support.","date":"2026-10-29","dateIsEstimated":false,"signalType":null,"signalTypeLabel":null,"sourceUrl":"https://www.assemblyai.com/blog/python-voice-agent-tool-calling","publisher":"assemblyai.com"},{"title":"Build a raw WebSocket voice agent: streaming STT, LLM and TTS with no SDK","summary":"AssemblyAI introduced raw WebSocket support for Universal-3.6 Pro Realtime and Text-to-Speech, allowing developers to build voice agents without SDKs. The update includes streaming STT, LLM Gateway integration, and TTS with barge-in and tool calling, all controllable at the protocol level. A 300-line Python example demonstrates the full architecture using three connections (STT, LLM, TTS) with one API key.","date":"2026-10-29","dateIsEstimated":false,"signalType":"feature_update","signalTypeLabel":"Feature","sourceUrl":"https://www.assemblyai.com/blog/raw-websocket-voice-agent-quickstart","publisher":"assemblyai.com"},{"title":"Build a production voice agent with LiveKit and AssemblyAI","summary":"AssemblyAI and LiveKit launched a production-ready integration enabling developers to build multi-user voice agents in ~100 lines of Python. LiveKit handles WebRTC transport and orchestration while AssemblyAI provides Universal-3.6 Pro Realtime for STT, LLM Gateway for reasoning, and Text-to-Speech API, all accessible via a single API key.","date":"2026-10-29","dateIsEstimated":false,"signalType":"integration","signalTypeLabel":"Integration","sourceUrl":"https://www.assemblyai.com/blog/build-production-voice-agent-livekit-assemblyai","publisher":"assemblyai.com"},{"title":"Best LLMs for Dictation Cleanup: Speed, Quality, and Cost, Benchmarked (2026)","summary":"AssemblyAI published a 2026 benchmark comparing over 50 LLMs on dictation cleanup tasks, emphasizing latency and exactness over raw quality due to real-time user expectations. The post introduces Voice Standard Bench’s scoring methodology and reveals that frontier models like Gemini 4 Argon lead in performance, while smaller models (e.g., Qwen3.5 4B Fast) excel in speed and cost efficiency. AssemblyAI also promotes its Dictation API as a fast-path alternative for transcription and cleanup in a single call.","date":"2026-10-28","dateIsEstimated":false,"signalType":"content_marketing","signalTypeLabel":"Content","sourceUrl":"https://www.assemblyai.com/blog/best-llms-dictation-cleanup","publisher":"assemblyai.com"}],"attribution":{"source":"Spyingbee","url":"https://spyingbee.com/updates/assemblyai","citation":"Spyingbee, \"AssemblyAI updates\", https://spyingbee.com/updates/assemblyai (retrieved 2026-11-22)"}}