TL;DR. Salesforce agreed to acquire Cresta, Five9 is buying Level AI for about $325 million, and SoundHound is cutting 27% of its staff. Soniox shipped STT v6, Hume launched an API that answers questions about audio, and Meta opened voice and avatar chats with Muse in the US. Apple opened Siri AI to all developers. US prosecutors charged three men over a $129 million AI voice-cloning fraud, and Elvis Presley's rights holders sued Suno.
New models and launches
Soniox released STT v6, which takes the voice agent's own text as input and filters the agent's voice out of transcripts (November 3). The new models, stt-rt-v6 for real-time streaming and stt-async-v6 for files, let developers send what the agent says over the same WebSocket as the user's audio, which Soniox says gives recognition context from both sides of the conversation. With agent text input enabled, agent voice echo cancellation keeps the agent's playback out of the transcript while still listening for the user, including when they interrupt. Soniox also reports better accuracy across 60+ languages, better speaker diarization, more reliable numbers, currencies and dates, and improved real-time translation; it published no benchmark figures or pricing in the announcement. The v5 models will be removed on December 15, 2026.
Sources: Soniox, Soniox docs
Reson8 released Rumble-2-Flash, a 0.5B-parameter speech-to-text model priced from €0.01 per hour of audio (November 3). Reson8 calls it its smallest and fastest speech-to-text model; it understands 40+ languages, 1,000 hours of audio comes to €10, and it is free to use until November 10. On AppTek's call-centre test, 302 hours of helpline calls in 14 English accents scored with AppTek's own script, Reson8 reports a 7.86% word error rate, and a 7.09% average across the seven languages it was tested on from the Hugging Face Open ASR Leaderboard (vendor figures).
Source: Reson8
Hume launched the Audio Decision API, which answers developers' own questions about a recording (November 2). Developers send audio with a question and a set of possible answers, such as yes or no, multiple choice, a rating scale or a time span, and the model returns a probability for each answer, without transcription or retraining; it can run dozens of questions at once on a live call or in batch over an audio library. Hume groups the questions into ten categories, including speech and language, emotion and expression, acoustic environment, conversation dynamics and communication quality. It reports 70.7% agreement with the average judgment of a human panel, against 56.5% for a single human rater, and an average score of 0.82 over 10 benchmarks at about 0.23 seconds median latency (vendor figures). Access is by waitlist, with a playground for testing; no pricing was published.
Sources: Hume, Hume product page
Meta opened voice conversations and avatar video chats with its Muse AI assistant to US users (November 5). Muse Voice lets users talk to Muse in real time, interrupt it mid-answer and pick from hundreds of voices. It runs on two models Meta first described in September: Muse Realtime Voice, which emits a stream of speech tokens that carry both the words and the delivery, and Muse Realtime Avatar, which reads the same token stream to animate the face and body so lip movement and expressions stay in sync with the voice. The avatar can be shown full screen, next to the chat or turned off for voice only. Meta's research post reports about 870 ms from user input to first response byte and 12 concurrent real-time sessions per GPU. The rollout covers mobile, web and Mac in the US; AI Glasses and the Muse Charm device are to follow, with no dates given.
Sources: Meta, Meta research blog
Google added Voice Remixing to Gemini 3.8 Flash TTS, which edits an existing voice from a text instruction (November 4). Developers start from a prebuilt voice, an Extended Voice Library voice or a custom voice, describe a change to age, pitch, energy, timbre or accent in plain language, and get back a new stored voice ID with a 24 kHz WAV preview. Remixed voices are created with gemini-3.8-flash-tts and can also be used for synthesis in Gemini 3.8 Flash-Lite TTS. It requires a paid-tier Gemini API project, with up to 200 stored voices per project, each kept for one year from last use. The feature is in the Gemini API and Google AI Studio.
Sources: Gemini API release notes, Gemini API docs
Underdog released Woof, an English TTS model that runs on a laptop CPU from a single file of about 32 to 41 MB (November 5). The model has 38.24M parameters (a 22.95M speech model and a 15.3M audio decoder) and runs with nothing but numpy, without a GPU, server or API key; a one-voice version is 31.7 MB and runs in the browser. It outputs 24 kHz audio in two voices, Juno and Mira, and includes a rule-based normalizer for phone numbers, dates, prices and codes. On Gradium's TTS Hard Cases benchmark the model card reports a word error rate of 1.42% for Mira and 1.53% for Juno, and 13.8 times faster than real time for Juno on one Apple M5 Max CPU thread (vendor figures). The weights are on Hugging Face under the Underdog Non-Commercial License 1.0: commercial use needs a paid licence.
Sources: Underdog Woof on Hugging Face, TechCrunch
Phonic launched Phonic-1, a speech-to-speech model for voice agents, and raised a $27.5 million Series A (October 29, missed last week). Phonic-1 processes and generates speech in one model instead of chaining transcription, a text model and speech synthesis. Phonic says it was trained on more than 50 languages with language switching mid-conversation, ranked first on the Artificial Analysis Speech to Speech Index at launch, and can run tool calls while the caller is still talking, shown in a 44-minute call with 123 tool calls (company figures). First Round led the round, with Lux Capital, Felicis and Definition, bringing total funding to $31.5 million. The model is available now.
Source: Phonic
Convoso made its no-code Voice AI Agent generally available for lead qualification and inbound calls (November 3). The agent runs inside Convoso's contact center platform, qualifies leads and warm-transfers sales-ready prospects to human agents. Convoso reports response times of about 600 milliseconds, more than 700 voices with tone and speed controls, optional voice cloning and language switching mid-call. In a 36-hour test, Tort Intake Professionals made more than 150 contacts, of which 10 to 12% were validated and transferred to human specialists (company figures).
Source: Convoso
Also shipped this week:
ElevenLabs announced native ElevenAgents integrations with NiCE CXone, Genesys Cloud, Amazon Connect and Five9, so companies can run its voice agents inside their existing contact center routing, with handoff to human agents that keeps the conversation context (November 4). Source: ElevenLabs
ElevenLabs also made SFX v3 (
eleven_text_to_sound_v3), its newest sound effects model, available to all workspaces; v2 will be deprecated on March 1, 2027 (November 2). Source: ElevenLabs changelogRime on-prem release
20261104cut Coda v2 time to first audio byte from 95 ms to 35 ms on A10G hardware (November 4). Source: Rime changelogSonos opened sign-ups for early access to Sonos 27voice, its new assistant for music, mood-based playlists, smart home control and general questions, for approved US users in English; wake-word detection runs on the speaker and requests go to Sonos's cloud. A full release is planned for 2027 (November 6). Sources: Sonos community, Engadget
Infrastructure and tooling
Deepgram opened Deepgram Marketplace, a catalog of 19 partner products for voice agents (November 4). The marketplace launches with partners in eight categories: voice models (Cartesia, Inworld), voice agent platforms (Vapi, LiveKit), testing and evaluation (Bluejay, Cekura, Coval, Hamming), audio enhancement (ai-coustics, Krisp), telephony and voice platforms (Twilio, jambonz, Vonage, UniSpeech), security and trust (Fortanix, SottoAI), avatars and video (LemonSlice, Tavus) and health signals (thymia). Eligible customers can apply their Deepgram commitment to partner products under one contract and billing statement. Each partner sets its own pricing and handles its own support.
Source: Deepgram blog
LiveKit Agents 1.8.11 added the Responses API to LiveKit Inference and speech translation models from OpenAI and Google (November 5). Developers can now call the Responses API with LiveKit Inference credentials, including structured output and web search, over REST or an optional WebSocket transport. New providers and models include Zoom Scribe streaming STT, Kalpa's kalpa-tts-1 (English and Hindi), Rime STT and the Prism speech-to-speech model, Fish Audio's transcribe-stream-1 realtime STT, OpenAI's gpt-realtime-translate and Google's gemini-3.5-live-translate-preview through a beta translation interface, and Respeecher TTS in Swedish, Brazilian Portuguese, German and Arabic. The release also sunsets Sarvam's bulbul:v2 TTS model.
Source: GitHub, LiveKit Agents 1.8.11
ElevenLabs made Eleven Realtime available to every ElevenAgents workspace (November 2). The rollout that began on October 19 is complete, and agents can pick Eleven Realtime v1 Mini, the faster default, or v1 Max, which reasons, as their voice mode; OpenAI Realtime clients can connect to its WebSocket. Speech to Text can now redact entities in realtime transcripts, sending a separate redacted_transcript event for each committed transcript. Evaluation criteria and data collection items can be managed through a new /v1/convai/analysis-items API, and ElevenLabs shipped the first release of a Unity SDK for voice agents (Unity 6.3 LTS or later).
Source: ElevenLabs changelog
Retell AI added OpenAI's GPT-Live-1 as a voice model for phone agents (November 2). The full-duplex model talks to the caller directly while the agent's existing text LLM keeps handling tools, knowledge bases and mid-call actions, so each agent has a voice prompt for conversation style and a backend prompt for logic. It works only for single-prompt voice agents on paid accounts, not for conversation flows, multi-prompt agents, transfer agents or chat agents.
Source: Retell AI changelog
Resemble AI now runs its deepfake detection natively inside Microsoft Teams (November 3). The Teams app analyzes call audio and video for synthetic media, identity mismatches and fraud signals and alerts the meeting host in real time, instead of joining the call as a separate participant, and produces a timestamped summary afterwards; users can optionally enroll face and voice samples linked to their Microsoft 365 accounts. It requires Microsoft 365 Business Standard, Business Premium, E3 or E5 (or equivalent A, G and Office 365 plans), because Business Basic, E1, F1 and F3 lack the compliance-recording APIs it uses. Pricing was not disclosed.
Source: Resemble AI
Also in the changelogs:
Vapi added nine GPT-Live voices in German, French, Hindi, Japanese and Korean, a
feedbackFilterEnabledsetting that drops caller speech repeating what the assistant just said from Deepgram transcripts, and API generation of simulation test suites; it also fixed mid-word silence gaps on GPT-Live phone calls (week of November 2). Source: Vapi changelogTwilio Voice API's Call resource now returns the STIR/SHAKEN attestation status of a call in a new
stir_verstatproperty (November 5). Source: Twilio changelogTelnyx added two-way local voice and SMS numbers in Estonia, India, France, Spain, the Bahamas and Japan, and says it now supports local outbound calling in 94 countries; its WebRTC Android SDK 3.8.0 adds automatic recovery from signaling and connection failures (November 3). Sources: Telnyx release notes, coverage, Telnyx release notes, Android SDK
Speechmatics realtime SaaS now ends a session when transcription falls more than 1 MB behind the audio, and improved recognition of some English terms (November 3). Source: Speechmatics changelog
The money
Salesforce signed a definitive agreement to acquire Cresta, which makes voice agents for contact centers (November 2). Cresta's software runs customer calls end to end with voice agents, coaches human representatives in real time and analyzes each conversation; Salesforce names Marriott, Comcast and Alaska Airlines as customers. Salesforce did not disclose the price. The Information reported $2.7 billion, citing a person familiar with the transaction. The deal is expected to close in the first quarter of Salesforce's fiscal year 2028, subject to regulatory approvals.
Source: Salesforce
Five9 agreed to buy Level AI for about $325 million in cash (November 5). Level AI makes AI applications for customer service conversations, runs on its own family of models called Level AI Latitude, and says it handles more than one billion customer interactions a year for customers including Gusto, Chime, Empyrean and VistaPrint. The two companies have worked together since 2022. Five9 expects the deal to close in 2026, and announced it alongside third-quarter revenue of $323.3 million, up 13% year over year.
Sources: Five9 (SEC filing), Five9 Q3 results (SEC filing)
SoundHound AI is cutting about 27% of its global workforce (November 5). In an SEC filing released with its third-quarter results, the company estimates $18 million to $22 million in pre-tax charges, mostly severance, with the cuts substantially complete by the end of 2026. Third-quarter revenue was $62.4 million, up 48.5% year over year, and SoundHound expects full-year 2026 revenue of $250 million to $260 million and adjusted EBITDA profitability in the second half of 2027. The company closed its acquisition of LivePerson on September 4.
Sources: SoundHound 8-K, SoundHound Q3 results
More deals:
Guava, a Los Angeles developer platform for voice agents, raised a $12.5 million seed round led by Acrew Capital, with Flourish Ventures, WillowTree Ventures, Clocktower Ventures and Commerce Ventures. It says it supports more than a dozen enterprise customers and tens of thousands of calls a day (November 3). Source: Business Wire
ElevenLabs CEO Mati Staniszewski said at Bloomberg Tech in London that starting preparation for an initial public offering in 2028 would be "a good goal." ElevenLabs was valued at $22 billion in an employee share sale in September (November 3). Source: The Next Web
Law, safety, and consent
US: prosecutors charged three men over a scheme that used AI-cloned executive voices to take about $129 million (November 5). The US Attorney's Office for the Southern District of New York announced an indictment against Jacob Soall, Sharon Hazan and Guy Sheetrit. Prosecutors allege the group used AI to mimic the voice, and sometimes the image, of a company's CEO on calls with staff, then posed as lawyers at international law firms to have money wired abroad and converted to cryptocurrency, between 2024 and 2026. Reported victims include hospital systems in Missouri and California, a UK fintech and a Rhode Island bank. Soall and Hazan were arrested in Spain and Sheetrit in the UK between June and September 2026, and the US is seeking extradition. The charges include wire fraud, money laundering and aggravated identity theft. These are allegations.
Sources: US Attorney's Office, SDNY, CBS News
US: Elvis Presley's rights holders sued Suno under Tennessee's ELVIS Act over imitations of his voice (November 6). Elvis Presley Enterprises, part of Authentic Brands Group, alleges that Suno trained on Presley recordings and lets users generate songs that imitate his voice, and that Elvis-like outputs could be prompted by his name as recently as October 2026. The suit is an early test of the 2024 Tennessee law on unauthorized use of a person's voice and likeness. The same legal team filed a separate copyright suit for Eminem's publisher Eight Mile Style. Suno says it "does not allow prompts naming specific artists or songs" and exists to help users make original music. These are allegations.
Sources: Music Times, Dork
Also this week:
Germany: Matti Klemm, the German dubbing voice of Jason Momoa and Mahershala Ali, plans to sue Netflix by the end of 2026, the voice actors' association VDS announced. VDS says a data access request showed Netflix had trained AI systems on Klemm's voice recordings without his knowledge, consent or payment; Netflix says its use of AI complies with applicable law (November 3). Source: Horizont
In the wild: products, enterprise, and culture
Apple opened Siri AI to all developers through App Intents (November 3). Starting with the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas, a Developer Preview lets any app bring its content and actions into Siri AI, so users can ask in their own words for a custom action such as checking a delivery, booking a reservation or placing an order. Apple also introduced "skills", which teach Siri AI to combine several actions into one request. Siri AI is available in English, French, Japanese, Korean, Portuguese and Spanish, with Traditional Chinese, Norwegian, Turkish and Vietnamese coming. Apple says the integrations will ship to customers "in a few months", with App Store submissions opening in the spring.
Source: Apple Developer
Apple Watch began beta testing Live Rewind and Siri Recap, which listen to conversations around the wearer (November 3). In the first watchOS 27.3 developer beta, Live Rewind shows a text snippet of the last 15 seconds of speech when the wearer double-presses the Digital Crown, and Siri Recap runs in the background to produce summaries of conversations through the day. Apple says audio is not recorded or stored; snippets disappear about 30 seconds after the display dims unless saved, and recaps can be set to delete after up to 7 days. The features need an Apple Watch Series 12 or Ultra 4 paired with an iPhone 16 or later, work in English only, are not available to users under 18, and will not be available in the EU or China at first.
Sources: Apple Support, TechCrunch
ElevenLabs now lets ElevenMusic users publish AI-generated tracks straight to SoundCloud (November 5). Users link a SoundCloud account, add a description, genre and tags to an original ElevenMusic track and click "Save to SoundCloud", without downloading and re-uploading the file. SoundCloud Artist Pro members get three months of ElevenMusic free, then $9.99 a month; the offer runs until October 31, 2027. TechCrunch reports that more than 25 million tracks have been generated with ElevenLabs' music models.
Sources: ElevenLabs, TechCrunch
SoundHound expanded its deal to put a voice co-pilot on Ultraviolette's electric motorcycles (November 4). VIOLETTE A.I. lets riders use a helmet headset to adjust traction control and ABS, run pre-ride checks on battery charge, range and tire health, query the owner's manual and handle navigation and calls. Ultraviolette exports to 20 countries and plans to add SoundHound's OASYS agent platform. The announcement came during EICMA 2026 in Milan.
Source: SoundHound via GlobeNewswire
Ambient scribe maker Heidi Health is cutting 18% of its staff as two cofounders leave (reported November 5). CEO Thomas Kelly announced the cuts on LinkedIn; Heidi has more than 600 employees. Kelly said "we weren't putting our investment into the right teams and markets to make the most of our continuing growth." Cofounder Waleed Mussa is leaving to spend time with family, and cofounder Yu Liu is moving to an advisory role until March 31, 2027. The cuts come about two months after Heidi announced a $100 million Series C led by Blackbird and a $240 million growth facility from General Catalyst's Customer Value Fund.
Sources: Startup Daily, HTN
Alexa+ launched in Japan in Early Access (October 29, missed last week). Customers in Japan can request access, with invitations sent on a rolling basis; after Early Access, Alexa+ will be free with Prime and cost ¥3,000 a month without it. Japan joins the U.S., UK, Canada, Mexico, India, Italy, Spain, Germany, Austria, Brazil, France and Australia, and Amazon says Alexa+ will reach more than 10 additional countries in 2027.
Source: Amazon
More voice agents went live:
WellNet Healthcare, a third-party administrator of health plans, is deploying EHVAhealth's voice agents on inbound provider calls for eligibility, benefits and claims status checks across medical, dental and vision plans, inside its existing Five9 contact center and Salesforce CRM (November 4). Source: EHVAhealth via PR Newswire
ElevenLabs says its own sales agent has handled more than 3,000 inbound leads this year in 32 languages and qualified over 90% of them without a person, according to field CTO Alex Holt (November 4; company figures). Source: Fortune
Apple changed its Siri AI support page so the assistant "will begin to download automatically" after setup, with a waitlist only "during times of high demand"; previously every user joined a waitlist (page dated November 5). Source: Apple Support
Papers this week
Full-duplex speech models can draft their next reply while still listening (arXiv, November 4). Researchers from The Chinese University of Hong Kong, Shenzhen and Amphion found that, across four chunk-interleaved full-duplex models, the next audio chunk leaves the output unchanged in 95.3% of chunks, and 88% of changes begin at the first token. Their training-free, lossless method, Speculative Listening, uses idle listening time to decode the next chunk in advance; on the eight-task TickTock-Bench it speeds up speak chunks by 2.6 to 2.8 times and tool-call chunks by up to 23.4 times, and cuts the decoding delay before a response starts by 73%, to 74 to 241 ms. Source: arXiv
InteractionBench scores GPT-Live, Gemini Live and open full-duplex models on timing, task accuracy, whispering and grounding (arXiv, November 2). AveraLabs ran 221 items per system across 13 scenarios in 1,519 graded runs. A turn-based cascade had the highest spoken-task score (38/50) and the second-lowest timing score, only Gemini Live whispered on most requests (8/10), and hosted systems passed 67 to 88% of groundedness probes against 14 to 35% for open full-duplex models. Timing is measured where the client hears the reply, so it includes network delay: removing about 1.5 seconds of delay for GPT-Live lowers its timing score from 86% to 56%. Source: arXiv
NetEase Youdao releases a streaming ASR model that never rewrites text it has already emitted (arXiv, November 4). Confucius4-R2T2 is Qwen3-ASR-1.7B post-trained on 1,500 hours of Chinese and English audio with the audio encoder frozen, aimed at voice agents and translators that act on each token as it appears; chunk size is configurable from 80 ms to 2 seconds. In the team's ranked comparison it has the lowest total average error (5.76) at 160 ms settings, including against three unnamed proprietary APIs, with lower mean emission lag than all three in both languages, though two of those APIs are more accurate in Chinese (vendor figures). Weights and code are on Hugging Face and GitHub. Source: arXiv
Multi-turn scores on transcripts do not predict how speech models handle the same conversation spoken (arXiv, November 3). Nankai University and Alibaba researchers built Omnilingua-AudioMC, 354 conversations and 1,404 rubric items in ten languages, aligned across transcript and speech input. Meta's Muse-Spark 1.2 had the highest transcript score (64.0) but dropped to 32.4 on direct audio, the weakest of six models, while Alibaba's Qwen3.8 Omni moved from 60.9 to 60.2 and led on speech; for the two models with the largest gaps, replacing earlier spoken turns with transcripts recovered more than replacing the current turn, in all ten languages. Source: arXiv
Runway details how it serves a 57B video model fast enough for live conversation with a voice agent (arXiv, November 3). Runway Characters runs an autoregressive video diffusion transformer of about 57 billion parameters at 704x1088 and 24 fps on four H100 GPUs. Caching and FP8 changes cut each diffusion step from 240.9 to 159.3 ms (147.4 ms with a fused kernel), and over seven weeks of production the worker averaged above 23.5 fps in 99.5% of sessions; the video pipeline takes 0.49 seconds of a 2.1-second median conversational turn (company figures). Source: arXiv
Listeners often hear a voice clone as a different person from the real speaker (arXiv, November 3). VIPBench, from Northwestern University and accepted to the NeurIPS 2026 Evaluations and Datasets track, collects 124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs of 100 public figures, including real recordings, TTS clones and voice morphs. In 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people; of eleven public speaker models tested, the best explains 61% of the explainable variance in listener judgments. Source: arXiv
Previous edition: Voice AI News, Week 44, October 26 to November 1, 2026