TL;DR. Deepgram is acquiring Daily, the company behind Pipecat, and the two published an open standard for passing tone between voice agent models. NVIDIA released Nemotron 3.5 Transcribe, AssemblyAI launched its first TTS model, Rime launched Prism and Cartesia opened Managed Agents. Broccoli raised $52 million. SAG-AFTRA voice actors ratified animation contracts with AI terms, and Amazon launched five new Alexa devices.

New models and launches

NVIDIA released Nemotron 3.5 Transcribe, an open-weight model for transcription and translation in 40 languages (October 28). The model, published as Nemotron-3.5-Transcribe-30B-A4B, is a Mamba-2 and Transformer mixture-of-experts model; NVIDIA's model card lists 34B parameters in total and about 4.3B active, including the audio encoder. It handles recordings up to one hour, labels up to eight speakers and translates between English and the 40 supported languages, with coverage that varies by pair. It works on recorded audio, not live streams. NVIDIA reports a 4.71% average English WER on the Open ASR Leaderboard sets and released an updated training dataset, Granary v2, alongside it. The weights are under the OpenMDW-1.1 licence.

Sources: NVIDIA on Hugging Face, TechCrunch

Rime launched Prism, a speech-to-speech model that writes each reply as text before speaking it (October 28). The text step gives companies a checkpoint they can inspect before audio goes out, and the model keeps listening while it speaks. Rime reports 0.57 seconds to first audio and a Ï„-Voice score of 56.6, which it says is 17 to 30 points ahead of OpenAI and Google realtime models that do not reason first (vendor figures). Rime calls Prism fine-tunable; the self-serve fine-tuning APIs are still being built with beta customers. Rime did not publish pricing or a language list.

Source: Rime

AssemblyAI released AssemblyAI-TTS-1, its first text-to-speech model, built for voice agents (October 27). It is priced at $15 per million characters with no concurrency fees or minimums, and covers English, Spanish, French, German, Italian and Portuguese at launch. AssemblyAI lists more than 4,000 voices plus instant voice cloning, 90 ms to streaming audio in production and a 47 ms median time to first byte in lab tests. It reports 99.0% word accuracy on standard text and 93.7% on hard pronunciation targets (vendor figures). The model is generally available.

Source: AssemblyAI

Cartesia launched Managed Agents, a hosted voice agent platform, at $0.03 a minute (October 29). Cartesia says the voice models and the agent harness were designed together for speech-to-speech conversations, and that Managed Agents ranks first of 30 agents on Artificial Analysis's Speech Agent Arena with a preference Elo of 1115. The $0.03 rate is launch pricing at 50% off, with no end date given. Agents can switch between English, Spanish, Hindi, French and Japanese mid-conversation, and Cartesia says 44 languages are coming.

Source: Cartesia

Deepgram made its Think-Live model and Flux TTS generally available (October 29). Think-Live decides when a voice agent should listen, wait or respond, and is generally available in the Deepgram Voice Agent API; a second model, Think-Fast, handles rules and tool calls and is in early access on the Deepgram LLM API. Flux TTS is now generally available in Spanish, Portuguese, French, German, Italian, Dutch, Russian, Hindi and Japanese, with 32 more languages in early access, and Flux STT Multilingual now covers 10 languages with 31 more in early access.

Source: Deepgram

Tencent open-sourced Hy-SimulMT, a Chinese-English simultaneous interpretation model (October 26). The 30B-A3B model translates speech in both directions while the speaker is still talking and can speak the translation in a clone of the speaker's voice. Tencent reports about 3 seconds of lag (StreamLAAL) on long-form speech and scores of 84.7 on RealSI-Sentence, 90.3 on FLEURS and 81.2 on ACL6060 (vendor figures). Weights, including FP8 and GGUF versions and an "Interleave" variant that outputs source and translation together, are on Hugging Face under Apache 2.0.

Sources: Tencent on Hugging Face, TechNode

Liquid AI released LFM3-Audio-2B and LFM3-Omni-2B, small speech models for laptops and edge devices (October 27). LFM3-Audio-2B takes text or audio and answers in text, speech or both; it transcribes, synthesizes and translates speech in 18 languages and turns spoken requests into tool calls, while its voice chat mode is English only for now. LFM3-Omni-2B (2.17B parameters) adds images and video, so it can answer spoken questions about a photo or a clip with its soundtrack. Both are part of the LFM3 family and use the LFM Open License v1.0, which requires a separate agreement for commercial use by companies with more than $10 million in annual revenue.

Sources: Liquid AI, Liquid AI on Hugging Face

Synthflow launched Vector, an AI agent for consumer lenders that follows an account across calls, texts and emails (October 28). Vector works on collections, loan applications, refinancing and servicing, keeps the account history between contacts and hands over to staff when a decision needs a person. In a four-week trial with a North American lender, Synthflow reports Vector reached borrowers 4.9% of the time against 3.5% for human teams, and that 49% of conversations ended in a payment commitment (vendor figures). It is available to lenders in the US, South America, the UK and the EU.

Source: VentureBeat

Maya Research released Maya 2 Native, an open-weight TTS model for 11 Indian languages and English (October 25, missed last week). The 2.5B-parameter model covers Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, with 20 built-in voices and zero-shot, cross-lingual voice cloning from 3 to 10 seconds of reference audio. It streams 24 kHz audio in 80 ms frames and ships with an OpenAI-compatible speech endpoint. Weights and inference code are under Apache 2.0.

Sources: Maya Research on Hugging Face, Analytics India Magazine

Also shipped this week:

  • Inworld updated Realtime TTS-2 with opt-in reasoning, in which the model plans how to deliver a line before speaking it, plus about 13% faster generation on short utterances and fewer skipped words in long passages. The model ID and price are unchanged (October 28). Source: Inworld

  • Filevine opened a beta of LOIS Voice Agents, which answer law firms' inbound calls, run intake with the firm's approved questions and file the recording, transcript and summary to the right lead (October 27). Source: Filevine via PR Newswire

  • Granola added dictation on Mac: pressing Fn turns speech into cleaned-up text in any app, on all plans; iOS and Windows are coming (October 29). Source: TechCrunch

  • Ollama added speech models to its Mac app, starting with Gradium's Phonon, which reads text aloud in five languages on Apple Silicon (October 24, missed last week). Source: Ollama

Infrastructure and tooling

Deepgram and Pipecat published the Real-time Context Standard, and Pipecat 1.15 implements it (October 29). RCS is an open format for passing context between the models in a voice agent beyond the words: a reading of how the user sounded goes at the head of the user's message, such as ([tone frustrated=0.86, neutral=0.10]), and short inline tags such as <<warm>>, [[laugh]] or [[pause 500]] tell the TTS how to speak the reply. Version 0.1 is an alpha and may change in ways that are not backward compatible; speaker identity and turn intent are on the roadmap. Pipecat 1.15, released the same day, adds "expressive conversation" (alpha) built on it, with tag adapters for Cartesia, ElevenLabs, Fish Audio, MiniMax, Inworld, Soniox, Mistral Voxtral 3, Gemini TTS and Azure DragonHD voices, and a Pipecat Tone model that reads emotion and sarcasm in each user turn on CPU. The release also adds an AssemblyAI streaming TTS service and a user turn strategy that lets a classifier decide when the user has finished. Deepgram announced RCS at its Deepgram Speak event, the same day it announced it is acquiring Pipecat maker Daily.

Sources: Pipecat docs, RCS 0.1, GitHub, Pipecat v1.15.0

Twilio made mid-call configuration and reconnect generally available in Conversation Relay (October 28). Apps can now change most session settings, including interruptibility, DTMF detection, timeouts, speech providers, models, voices and hints, over the live WebSocket without dropping the call. With applicationReconnect="true", Twilio buffers up to 500 events during an unexpected connection drop and replays them when the app reconnects. A new echo prompt filtering option stops the agent from reacting to its own synthesized speech leaking into the caller's microphone, and Conversation Relay now supports ElevenLabs v3 Conversational and v4 Turbo voices, which use natural-language audio tags instead of SSML.

Source: Twilio changelog

Telnyx now hosts Pipecat voice agents on its Edge compute platform (October 28). Developers deploy an unmodified Pipecat agent with telnyx-edge ship and connect it to a Telnyx phone number, with no servers or containers to run. The phone number, call audio, Flux speech-to-text, LLM inference and text-to-speech all run on Telnyx under one API key. Agents run in every Telnyx Edge location, on Python 3.11 with WebSockets as the default runtime.

Source: Telnyx release notes

LiveKit Agents 1.8.10 added Reson8 speech-to-text and Mistral Voxtral 3 models (October 29). The new Reson8 plugin does batch and realtime transcription with server-side turn detection. The Mistral plugin adds the Voxtral 3 transcription, realtime and TTS models with 36 TTS voice presets, and drops Chat Completions for Mistral's Conversations API, a breaking change. Inworld's inworld-tts-2-flash and AssemblyAI's assemblyai-tts-1 are available through LiveKit Inference. Streaming STT now detects stalled connections so fallback can take over, and RecorderIO can record at a constant bit rate. The JavaScript SDK (1.9.5, October 28) added DecisionModel, AssemblyAI TTS on LiveKit Inference and language_hint for Deepgram Nova-3.

Sources: GitHub, Python 1.8.10, GitHub, JS 1.9.5

ElevenLabs added a fast LLM tier for agents and clone strength for v4 voices (October 26). ElevenAgents can run supported LLMs on a fast service tier, which costs more per token for higher throughput and lower latency; standard stays the default. Text to Speech and Text to Dialogue accept clone_strength, from 0 to 1, which sets how closely a Professional Voice Clone tuned for a v4 model follows its training samples. Realtime Speech to Text now accepts silence thresholds down to about 0.19 seconds, SIP trunk calls forward the User-To-User header to the conversation initiation webhook, and call billing now stops when the call ends. Service account API keys can be set to expire after 15 minutes to 90 days.

Source: ElevenLabs changelog

Microsoft released LLM Speech 09-2026 for Azure AI Speech (October 26). The update to the Fast Transcription API adds Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu, Catalan, Malay and Welsh, for 50 languages across 60+ locales, and transcribes mixed-language audio without a locale hint. Microsoft reports up to 2x faster processing than the previous 2607 release and, on internal evaluations, up to 25% better token error rate and 37% better entity error rate. Requests accept up to 2,000 phrases in a phrase list. The update deploys automatically.

Source: Microsoft Foundry Blog

Changelog round-up:

  • Deepgram stopped supporting the dg-token header for verifying callback requests and says to switch to Basic Auth, and launched a new U.S. endpoint (October 27). Source: Deepgram changelog

  • Vapi added OpenAI's gpt-transcribe as a transcriber with keywords and a prompt, allowed AssemblyAI on HIPAA-enabled assistants, added a per-assistant Traffic tab showing how calls split across published versions, and added regional OpenAI models (week of October 26). Source: Vapi changelog

  • Speechmatics made Oak 1, its medical model, available for real-time transcription with model set to oak-1, and stopped inserting spaces between words in Chinese and Japanese transcripts (October 27). Source: Speechmatics changelog

  • Rime on-prem release 20261028 moves Coda v2 to a new checkpoint, and Rime no longer publishes new Coda v1 image tags (October 28). Source: Rime changelog

  • Hugging Face Transformers 5.21 added Gradium's Phonon TTS, Kyutai's Pocket TTS and the Pocket TTS audio tokenizer (October 26). Source: GitHub, Transformers v5.21

  • Synthflow now lets users compare any two versions of an agent, including the draft, instead of only consecutive versions (October 30). Source: Synthflow changelog

The money

Deepgram is acquiring Daily, the company behind the open-source Pipecat voice agent framework (October 29). Daily's team joins Deepgram, and Daily co-founder Kwindla Hultman Kramer becomes Deepgram's Head of Open Source. Deepgram says Pipecat will stay free and open source under its existing BSD license, and that an independent Pipecat Foundation is being set up with contributing organizations. The price was not disclosed; TechCrunch reports that Daily had raised more than $62 million, citing PitchBook.

Sources: Deepgram, TechCrunch

Broccoli raised $52 million for AI agents that answer calls for home-service businesses (October 28). The total combines a $22 million Series A and a $30 million Series B raised nine months apart, both led by Khosla Ventures, with Y Combinator and Operator Partners taking part. Broccoli's agents answer calls, book appointments, follow up on estimates and renew service memberships for plumbing, HVAC and electrical companies; it reports more than 1,500 business customers, over 10 million customer conversations and 2.5 million appointments booked (vendor figures). Fortune, citing a person familiar with the matter, puts the post-money valuation at $300 million; Broccoli declined to comment.

Sources: Broccoli release, Fortune

More deals:

  • Twilio reported third-quarter revenue of $1.56 billion, up 20% year over year (October 29). On the earnings call, management said voice revenue grew more than 20%, voice revenue in the self-serve channel grew 55%, and voice and consumer AI account for about 1% of total voice interactions; it also cited an eight-figure deal with a health insurance provider using ConversationRelay. Sources: Twilio results, earnings call transcript

Law, safety, and consent

SAG-AFTRA members ratified the TV animation contracts, with new AI terms for voice actors (October 29). Members approved the 2026 Television Animation Agreement and the 2026 Basic Cable Animation Agreement with the AMPTP by 94.95% to 5.05% after voting closed on October 29. The four-year contracts run from July 1, 2026 to June 30, 2030 and include a 12.6% wage increase over the term. According to the union's presentation, as reported by Voice Over Herald, performers must specifically consent before their digital replica is used to dub a performance into another language, producers agreed that a synthetic performer needs "significant additional value" to be used in a covered role, and studios must notify the union if performances are licensed to train public commercial AI systems.

Sources: AWN, Voice Over Herald

A Los Angeles mayoral candidate filed complaints over a phone line that uses an AI clone of her voice (October 26). Nithya Raman's campaign filed complaints with the California Fair Political Practices Commission and the Los Angeles City Ethics Commission against Thrive LA PAC, which backs Mayor Karen Bass. Signs advertising "Nithya Raman Realty" direct voters to a number where an AI voice speaks in the first person as Raman and answers questions; the campaign cites California Elections Code Section 20012 on fabricated candidate audio close to an election and Government Code Section 84514 on AI disclosure in political ads. Thrive LA's executive director Ana Dahan said: "AI was a tool to express our First Amendment right for political satire." These are allegations.

Sources: FOX 11 Los Angeles, Hoodline

A super PAC is using an interactive AI voice in robocalls against Texas Senate candidate James Talarico (reported October 29). The calls from Lone Star Liberty, which supports Republican Ken Paxton, use an AI voice called "Jennifer" that introduces itself as a virtual staffer and responds to voters in real time, attacking Talarico over alleged data center money. Call-blocking service Nomorobo, which helped CNN investigate, says it first detected the campaign on October 23 and that it came from dozens of numbers and reached a relatively small number of Texas landlines. Unlike the Raman case, the voice is disclosed as AI and does not imitate a real person.

Source: Nomorobo

Also this week:

  • FCC published its revised rules on revoking consent to robocalls, which take effect November 25 (October 26). Callers may designate one exclusive way to revoke consent and treat a revocation as covering only the category of informational calls it targets; the FCC's February 2024 ruling counts AI-generated voices as artificial voices under the TCPA, so the rules cover AI voice agent calls. A companion proposal on shorter revocation deadlines and a "revoke all" method takes comments until November 25. Sources: Federal Register, rule, Federal Register, proposal

  • Greene v. Google: Judge Shella Deen made her ruling final, overruling Google's demurrer on all four claims in former NPR host David Greene's suit over NotebookLM's male AI host voice in an order filed October 23 (missed last week). She wrote that whether Google built the voice from Greene's recordings "or from another source that merely resembles him is a factual question." These are allegations, and Google denies them. Sources: Radio Ink, Hoodline

  • Equity, the UK performers' union, told members to "proceed with extreme caution" over a video game casting whose contract signs away all rights to a performer's voice and likeness, including for AI training, for a single buyout fee (October 23, missed last week). Equity said a non-disclosure agreement stopped it from naming the game or studio. Source: Voice Over Herald

In the wild: products, enterprise, and culture

Amazon launched five Alexa devices, including an Echo with a camera that Alexa+ can see through (October 28). The new Echo Dot ($99.99) adds a built-in smart home hub, an eero Wi-Fi extender and Amazon's AZ3 chip with an AI accelerator that Amazon says sharpens conversational detection and filters background noise. Echo Studio with Home Camera ($249.99) lets Alexa respond to what it sees, for example explaining the rules on a board game box held up to it or recognizing a cereal box to reorder it. The lineup also includes the Echo Time bedside clock ($199.99), the Alexa Photo Frame (from $199.99) and a redesigned Echo Show 16 (from $499.99, preorder). Amazon says hundreds of millions of customers are using the new Alexa experiences and talk to Alexa two to three times more; Alexa+ is included with Prime. These are company figures.

Sources: Amazon, TechCrunch

Apple's Siri AI added five languages in beta and a memory of past conversations (October 26). With iOS 27.2 and iPadOS 27.2, Siri AI runs in beta in French, Japanese, Korean, Portuguese and Spanish as well as English. TechCrunch reports Siri can now remember details from past conversations and draw on photos attached in Messages, calendar attendees, voicemails and Safari history, and users can exclude specific apps. The same day, Apple announced a new iPad mini with the A20 Pro chip, which Apple says supports more expressive Siri AI voices and more accurate systemwide dictation.

Sources: Apple, TechCrunch

Rivian is licensing parts of its voice assistant to Volkswagen (October 29). In its third-quarter earnings materials, Rivian said it expanded its joint venture with Volkswagen Group by "licensing elements of our AI Voice Assistant technology as part of the broader electrical architecture and software partnership." Rivian Assistant reached the R2 in October after launching on the R1 models. Rivian did not disclose fees or which Volkswagen models will use it.

Sources: Rivian (SEC filing), TechCrunch

Concentrix now translates live voice calls between customers and advisors in more than 60 languages (October 27). With Live Voice Translation, part of its iX Suite, each side of the call hears the conversation in their own language. Concentrix gives examples of advisors who speak Southeast Asian languages supporting English-speaking customers, and of advisors switching to Mandarin when a customer preferred Chinese. It reports "97%+ productivity levels" in real-world conversations without defining the measure. These are company figures.

Source: Concentrix

An Australian credit union's voice agent now takes 7% of its calls (October 27). Southern Cross Credit Union built Nova on Salesforce Agentforce Voice in six weeks, after wait times rose from under two minutes to almost eight. Nova gives balances to 95% of callers who choose the balance option and pass verification, and hands frustrated callers to staff with the conversation context. It runs alongside the credit union's Vonage contact center. These are Salesforce and SCCU figures.

Source: Salesforce

A hijacked podcast kept publishing with an AI clone of its host's voice (October 22, missed last week). Sanderson Jones's show Lifefulness: Live Life Fully stopped in 2021. In May 2026 new episodes appeared that he did not make, voiced by a clone of his voice, after his feed was copied and moved to a different host. The hijacked feed earned $1,293 on RSS.com, which RSS.com paid to Jones before helping return the show. The case was told on the Hyperfixed podcast on October 22.

Sources: Podnews, Voice Over Herald

More deployments and products:

  • Mazda began rolling out Google Gemini by over-the-air update to the new CX-5 with Google built-in in Europe, replacing Google Assistant once drivers enable it, in 12 languages, with three years of free data (October 26). Source: Mazda UK

  • Pypestream moved its Voice AI from beta to general availability and said quick-service restaurant brands run it in production; one brand's national guest line went from brief to live in four weeks. No brands were named (October 28). Source: Pypestream (Yahoo Finance)

  • Expedia Group added an AI Voice Agent for Vrbo customer support that handles common requests and passes the call context to a human agent when needed, one of more than 30 features in its fall release (October 28). Source: Expedia Group (Hospitality Net)

  • Leeds Teaching Hospitals NHS Trust will roll out TORTUS ambient voice note-taking across the Leeds General Infirmary emergency department this autumn, after a February to April pilot cut documentation time by 61%, about 13 minutes per consultation (October 28). Source: National Health Executive

  • TELUS Digital published results for its ElevenLabs outbound agent that welcomes new TELUS home internet customers: across about 2,300 calls in the proof of concept, 30-day cancellations were less than half the rate of the wider new-customer group, satisfaction averaged 8.5 out of 10, and the agent went from vendor selection to first call in about six weeks (October 30). Source: ElevenLabs

  • Voicemod signed a licensing deal with Disney for Star Wars and Marvel soundboards, starting with The Mandalorian and Grogu at $9.99. The packs are prerecorded clips, not AI recreations of the characters' voices (October 30). Source: GamesBeat

  • Providence Public Schools parents asked for notice and an opt-out from AMIRA, an AI reading program that listens to K-5 students read aloud, and asked what happened to recordings made before the district changed its settings in August. The district says the system does not keep new voice recordings (reported October 28). Source: WJAR Turnto10

  • TK Elevator service technicians use a voice agent built with Microsoft Foundry and Azure OpenAI that asks them about each job afterwards and records their account in their own words; TKE has 25,000 technicians (October 22, missed last week). Source: Microsoft

Papers this week

  • ServiceNow details Apriel Voice 27B Preview, an open speech model that calls tools directly (arXiv, October 29). Built on Google's Gemma 4 26B-A4B checkpoint, the model processes speech and produces both replies and tool calls without a separate dialogue model, while keeping text and image input. In thinking mode the team reports 98.90 on Big Bench Audio, 81.13 on Speech-BFCL and 65.51% task success on Ï„-Voice with a mean latency of 1.91 seconds (vendor figures); weights are on Hugging Face. Source: arXiv

  • Spoken agents can start a tool call before the user finishes talking (arXiv, October 26). Nanyang Technological University researchers built 26,694 synthetic dialogues (786 hours, 48 domains) labelled with the earliest point at which a request becomes actionable, and trained an omni-modal model to decide chunk by chunk. It reaches tool-state macro-F1 of 0.48 on held-out dialogues and 0.47 on an audio version of τ²-bench; for correct calls, the estimated lead over waiting for the end of the turn is 1.8 to 2.0 seconds in-domain and 1.1 seconds on τ²-bench. Source: arXiv

  • End-of-turn detectors trained on real conversations cut false alarms on mid-turn pauses (arXiv, October 29). Uniphore researchers trained the same streaming detector on matched human-human conversational and single-turn human-machine corpora; the conversational model scored up to 0.45 higher F1 across four test sets, a gain that tests on single-turn speech alone do not show, and their 1.8M-parameter model beat open-source baselines on every test set. A companion paper finds that one small head on a frozen multilingual encoder transfers across five CALLHOME languages, with a few hours of target-language audio recovering nearly all the loss. Sources: arXiv, arXiv

  • VoxLock filters background talkers out of voice-agent calls without an enrollment step (arXiv, October 29). The open model, a fine-tune of the 27M-parameter REAL-TSE extractor, takes its voice sample from the call itself; on Krisp's 265-recording benchmark it brings WER to 4.95% from 23.21% unprocessed, not significantly different from Krisp's best filters, and on 400 AMI meeting windows it cuts WER from 42.95% to 10.63%. The 8-bit model runs a live call on 77% of one CPU thread with 24 ms of algorithmic latency (single-author results; model, code and per-file scores are released). Source: arXiv

  • Leaked speech tokens are enough to clone a user's voice saying new text (arXiv, October 26). Harbin Institute of Technology researchers show that the representations an on-device speech front end sends to a server still carry speaker identity: from a leaked Moshi representation covering 3 seconds of speech, their attack generates attacker-chosen text with an average CAM++ speaker similarity of 0.78 and a target-text WER of 3.80%, without the original audio or its transcript. Source: arXiv

  • Local ASR-LLM-TTS pipelines take 1.28 to 1.54 seconds to first audio in a 999-question test (Electronics, October 30). Obuda University researchers ran 999 read-aloud HeySQuAD questions through local pipelines; on a validation subset with streaming Vosk, median time from end of speech to first audio was 1.28 to 1.54 seconds (2.03 seconds on a 27B configuration), and endpoint delay accounted for most of the wait except in that 27B case. Vosk's corpus WER was 0.200 against 0.049 for buffered Whisper, 4-bit quantization cut device memory by about two to three times, and a majority of answers passed the full spoken-format checks only with 27B to 32B models. Source: MDPI Electronics

  • Google releases a dataset for audio descriptions that must fit a video's silences (arXiv, October 29). MTG-AD, accepted to the NeurIPS 2026 Evaluations and Datasets track, holds 36 hours of user-generated video with more than 8,000 descriptions co-written by blind and sighted professionals, plus a reference-free evaluation that renders scripts to audio and checks overlap with speech. Frontier multimodal models described visual events accurately and reduced speech overlaps but showed gaps in pacing and flow. In ratings that included seven blind and low-vision participants, users narrowly preferred human-written descriptions but found model-generated ones highly beneficial for comprehension. Source: arXiv

  • A streaming English ASR model runs in real time on an ESP32 microcontroller (arXiv, October 28). A University of Science and Technology Houari Boumediene researcher quantized the 65.7M-parameter Kroko transducer to 2 and 4 bits without retraining, shrinking 71 MB of int8 weights to 26.65 MB, and an ESP32-P4 with no neural accelerator processes each 1.28-second audio chunk in 0.90 seconds. LibriSpeech WER is 4.34% on test-clean and 10.26% on test-other, against 3.94% and 9.17% for the original model; energy was not measured. Source: arXiv

Previous edition: Voice AI News, Week 43, October 19 to 25, 2026