TL;DR. Mistral launched Voxtral 3 with a voice agent API, ElevenLabs released Eleven Realtime, its first speech-to-speech models, and Block previewed Dessa, a full-duplex model trained only on audio. IBM released Granite Speech 5.0. Otter.ai secured $320 million from General Catalyst's Customer Value Fund and Dayforce agreed to buy Ribbon AI. A judge signalled that a former NPR host's voice suit against Google can go ahead, and Wyndham's AI concierge now answers calls at more than 1,700 hotels.
New models and launches
Mistral launched Voxtral 3, speech models plus a voice agent API (October 20). The family covers Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3 (4B parameters), Voxtral Mini Transcribe Realtime 3 and Voxtral Mini TTS 3. Transcription covers 50+ languages (over 35 in real time, under 500 ms), and TTS covers 35+ languages and 50+ dialects. A voice agent SDK built on LiveKit chains transcription, a reasoning model with tool calls and speech in one streaming pipeline. Prices start at $0.0035 a minute for batch transcription, $0.008 for real time, $13 per million TTS characters and $0.04 a minute for voice agents on Mistral Small 4. The models run on Mistral's EU-hosted API, US and global endpoints, or self-hosted.
Sources: Mistral AI, The Decoder
ElevenLabs released Eleven Realtime, its first speech-to-speech models for voice agents (October 20). Eleven Realtime v1 Mini targets low-cost conversation and v1 Max targets multi-step tasks with tools. ElevenLabs reports, using Artificial Analysis evaluations, 98.8% on Big Bench Audio for Max and 75.9% for Mini, with average time to first audio of 0.57 and 0.62 seconds. Both support 90+ languages and the 17,000+ voices in its library, and the API is compatible with OpenAI's Realtime API. A day later ElevenLabs put Eleven Realtime Translation, live voice translation in 90+ languages built on Eleven v4 Turbo, into alpha for enterprise API customers.
Sources: ElevenLabs, ElevenLabs (Realtime Translation)
Block's AI lab previewed Dessa, a full-duplex speech model trained only on audio (October 23). Dessa listens and talks at the same time and was trained from scratch on recordings of human conversation, with no text and no pretrained models. Block says it was trained on one NVIDIA GB200 NVL72 system and runs faster than real time on a MacBook Pro with an M4 Max chip or better. It is a research preview: there is a public waitlist and early access for a small group of AI safety researchers, with no API, pricing or weights.
Sources: Block, TechCrunch
IBM released Granite Speech 5.0, led by an open 2B transcription and translation model (October 21). Granite Speech 5.0 2B transcribes six languages and translates across eight, and IBM reports a 5.06% average English WER, second on the Open ASR Leaderboard. It was trained on 390,000 hours of public audio plus synthetic data. IBM also made its 470M Turbo CTC model multilingual at 19,870 RTFx and updated Granite Speech 4.1 2B Plus for multi-speaker recordings. All three are under Apache 2.0.
Source: IBM Granite on Hugging Face
Salesforce AI Research introduced Launa-Omni, a real-time voice model for enterprise work (October 19). Launa-Omni pairs an audio-adapted text model with a separate in-house TTS model, so the text of each reply is visible before it is spoken. Salesforce reports 59.69% on Audio MultiChallenge, 90.60% on VoiceBench and a 72.9% macro score on the τ³-Voice customer service benchmark. A simplified version runs in Slackbot Voice Dictation; Salesforce gave no wider release date.
Source: Salesforce
Gradium open-sourced Phonon, on-device TTS models from 64 MB (October 20). Phonon 31m (64 MB) is English-only under Apache 2.0, and Phonon 40m (74.6 MB) comes as five single-language models and one multilingual model covering English, French, German, Spanish and Portuguese under CC BY-NC. Gradium reports a 1.70% WER on structured text for Phonon 31m against 3.50% for Kokoro, and 7 ms to first speech on a MacBook Air M5 CPU. The models run on CPU, GPU and NPU, including in the browser over WebGPU. Gradium presented the release at NVIDIA GTC Berlin on October 21.
Sources: Gradium, Gradium (announcement)
Chariot launched Chariot-1-ASR Realtime for 10 Indian languages and Indian English (October 22). Across the 10 languages, Chariot reports 9% fewer errors than Sarvam Saaras v4, 20% fewer than ElevenLabs Scribe v2 and 26% fewer than Google Gemini 3.5 Transcribe, all batch systems; against streaming systems the gaps are 17%, 57% (Scribe v2 Realtime) and 34% (Gemini 3.5 Transcribe Live). On Hindi phone calls it reports 5.04% WER against 8.27% for OpenAI gpt-transcribe. The tests use 79 public test sets, and the model is live on the Chariot platform in balanced and high-accuracy modes.
Sources: Chariot, Analytics India Magazine
Scaled Cognition released Voice, a TTS model for customer service at $9 per million characters (October 20). Self-hosting costs $1 per million characters before infrastructure. The company lists 70 ms from text to speech, more than 25 languages and voice cloning, and says the model reads abbreviations, dates, addresses and account numbers from context. Voice is generally available and connects to existing agents, contact center platforms and IVR systems.
Sources: Scaled Cognition via CMSWire, SiliconANGLE
Amazon upgraded Polly's Generative engine and added 14 voices (October 22). The upgrade is a drop-in replacement with no API change or price change, aimed at expressiveness and at reading numbers, currency and addresses. The 14 new voices span nine locales, Hindi and Finnish became generally available, and the catalog now has 57 generative voices in 11 languages and 25 locales. It runs in 11 AWS Regions, now including AWS GovCloud (US-West).
Source: AWS
monday.com launched Harmony, a voice AI platform for sales calls (October 19). Harmony agents prospect, qualify leads, handle objections and transfer warm calls to staff, using task-specific in-house models on voice technology monday.com acquired from OneAI, built on Twilio. The company reports about 200 ms from audio in to audio out. It is sold to mid-market and enterprise teams in the US, UK, Australia and Singapore; client Bright Car Finance reports a 9% increase in closed deals in less than a year.
Source: monday.com
Also shipped this week:
VoxNexus, based in Singapore, launched VoxEdge, an on-device Android voice stack (recognition in 36 languages, translation in 103, synthesis in 45), and VoxSense, always-on listening that it says uses 162 mAh over 12 hours (October 23). Source: VoxNexus via TradingView (EQS)
Infrastructure and tooling
Zoom opened voice agent, text-to-speech and interpreter APIs to developers (October 20). At Zoomtopia, Zoom added 15 services to Zoom AI Services. Voice Agent APIs run speech recognition, LLM reasoning and speech synthesis as one managed pipeline, new TTS APIs cover synchronous and live streaming, and an Interpreter API chains transcription, translation and synthesis for near-real-time speech-to-speech interpretation. The Scribe transcription API gains automatic language detection, time-synchronized captions and new "3B Pro High-Accuracy" models. A day earlier Zoom published Decisions and Decisions Pro, which score text and images against developer-defined outcomes, at $0.042 and $0.10 per million input tokens; Zoom reports 81.3% and 83.9% average accuracy on five text benchmarks (vendor figures).
Sources: Zoom, Zoom AI Research
Sierra released τ-bench 2.0 with voice agent tests in six languages (October 19). The open-source benchmark adds τ-i18n, which runs the same airline, retail and telecom tasks in English, Spanish, Brazilian Portuguese, Hindi, Korean and Mandarin, in text and full-duplex voice, plus τ-entity, 200 outbound-call tasks that test exact capture of names and numbers. In Sierra's write-up (October 21), across three voice agents and 2,700 calls, Spanish, Hindi and Portuguese trailed English on task completion by 2.4, 3.8 and 6.4 percentage points; Korean and Mandarin scored lower. GPT Live 1 and Gemini 3.8 Live missed a caller turn, replied slowly or failed to yield in more than 80% of calls, against 16% for Grok Voice Think Fast 2.0.
Sources: GitHub, τ-bench 2.0.0, Sierra
Pipecat 1.14 added Zoom speech-to-text and NVIDIA Nemotron Omni services (October 24). ZoomAISTTService streams transcription through Zoom Scribe Live, with optional speaker labels. NvidiaOmniLLMService lets a Nemotron Omni model hear the user directly and replace the STT step. Speechmatics STT now accepts 8 kHz telephone audio, the Twilio serializer can authenticate hang-ups with an API key instead of the account Auth Token, and the release adds a Perplexity Agent API service and Smallest AI LLM and STT options. Version 1.13 (October 19) introduced LLMWithBackend in alpha, which pairs a conversational front-end model with a back-end model that handles slower reasoning and tool calls.
Sources: GitHub, v1.14.0, GitHub, v1.13.0
LiveKit Agents 1.8.9 runs decision models through LiveKit Inference (October 24). inference.DecisionModel evaluates named decisions, such as whether a caller wants a human, with LiveKit credentials instead of a provider key; access requires a project feature flag. New providers include a ThunderPhone speech-to-speech plugin, AssemblyAI's assemblyai-tts-1 streaming TTS with word timestamps in six languages, Fish Audio Drama 3 and Speechify simba-3.5. Background tools can now report progress without triggering a spoken reply. The Rime WebSocket plugin now requires an explicit model. The same day, the JavaScript SDK (1.9.4) added MCP server integration.
Sources: GitHub, Python 1.8.9, GitHub, JS 1.9.4
Deepgram expanded Nova-3 multilingual to 55 languages in one model (October 23). Requests with model=nova-3&language=multi now reach a model that transcribes 55 languages with code-switching and per-word language tags, on the same batch and streaming endpoints. A new optional language_hint parameter biases detection toward expected languages, and responses echo the normalized hints in languages_hinted. The previous version can still be pinned. On October 20 Deepgram shipped its October self-hosted release, which requires NVIDIA driver 580 or later.
Source: Deepgram changelog
Rime added speech-to-text to its API (October 20). The TTS company now streams English and Spanish recognition over WebSocket or gRPC, with partial transcripts while the person speaks. It offers a written mode for formatted text and a word-for-word mode that keeps every spoken word, plus context terms for names and phrases. It uses the same API key as Rime's TTS and has Python, Node.js, Go and Rust SDKs. Rime did not list pricing or latency in the entry.
Source: Rime changelog
ElevenLabs added alerts to ElevenAgents and made guardrails generally available (October 19). Alerts flag a workspace when conversation failures, tool failures, webhook errors, guardrail triggers or call initiation errors pass a threshold, and can notify by webhook, Slack or PagerDuty. Speech Engine conversations can now run up to 40 minutes, up from a 10-minute default. The new Realtime Translation WebSocket API detects the source language and has an eagerness setting that trades speed for quality. The Eleven Realtime agent models accept existing OpenAI Realtime API clients after an endpoint and session config change.
Source: ElevenLabs changelog
Changelog round-up:
OpenAI made its Decisions API generally available with
gpt-6-luna, at up to 400 questions per request, up from 200 in beta (October 22). On October 20 it added 12-hour prompt caching, with cache writes at 2.5x the input token price. Source: OpenAI changelogVapi added a priority processing toggle that runs supported OpenAI models in OpenAI's Fast tier, alerts for failed server webhooks, and Claude Haiku 5.5, Gemini 3.8 Flash and Cartesia Sonic-3.6 (week of October 19). Source: Vapi changelog
Speechmatics brought the updated Melia 1 model, with a 1.6% English word error rate improvement, to real-time transcription, and
output_localenow accepts a list of locales (October 19). Source: Speechmatics changelogTelnyx voice AI assistants can now send AssemblyAI Universal-3.5 Pro a context prompt of up to 1,750 characters and up to 100 keyterms, and pass the agent's last reply on every turn (October 20). Source: Telnyx
Twilio made account-wide default call recording settings generally available in its US1 region, with au1 and ie1 planned (October 20). Source: Twilio changelog
SIPPIO opened an AI developer program with a sandbox for connecting voice agents to its phone network in more than 80 markets, following AI Connect for Microsoft Foundry agents (October 20). Source: SIPPIO
Toku launched Makimoto Flows, a visual builder that chains transcription, summary and call-tagging blocks and exports them as Python, with 1,000 free minutes a month and processing in Singapore (October 21). Source: PR Newswire
The money
Otter.ai closed $320 million in growth financing from General Catalyst's Customer Value Fund (October 20). Otter says the money goes to enterprise growth, go-to-market, international expansion and agentic workflows built on meeting conversations. TechCrunch describes the Customer Value Fund as non-dilutive capital that earns its return on the results of marketing spend, not an equity round, and no valuation was given. Otter reports 40 million users, use by teams at 86% of the Fortune 500 and more than 1 billion meetings transcribed (vendor figures); it crossed $100 million in ARR last year and had raised $50 million before this deal, most recently in 2021.
Sources: Otter.ai, TechCrunch
Dayforce agreed to buy Ribbon AI, which runs job interviews with voice agents (October 23). Ribbon's conversational voice and video AI screens candidates and also handles sourcing and outreach; it reports hundreds of customers and more than one million completed interviews. After closing, the technology will be built into Dayforce's new Agentic Recruiting product. Price and closing date were not disclosed.
Source: Dayforce
Nirva raised $25 million for a pendant that listens only to its wearer (October 22). The pre-Series A round brings Nirva's total to $33 million, after $8 million earlier this year from investors including Lightspeed and South Park Commons. Users record voice samples at setup and the device filters out other people's speech; by default it keeps no audio or transcripts after processing. The 8-gram pendant costs $249, lasts up to three days per charge, opens for orders in November and ships at the end of 2026.
Source: TechCrunch
More deals:
Karta AI, a Bengaluru startup building voice agents for inbound calls at banks and lenders, raised a $3 million seed round co-led by PeerCapital and Cedar Hill Capital, with Wyser VC, CDM Capital and Seedstars (October 22). It reports about two million customer conversations a month in more than 25 languages. Source: CIOL
CNTXT AI, an Abu Dhabi data and AI company whose products include the Munsit Arabic speech-to-text and text-to-speech models for more than 25 dialects, said it has more than $300 million in committed revenues, mostly from US hyperscalers, AI developers and large enterprises, and opened its first US office in the San Francisco Bay Area (October 19). Source: CNTXT AI release
Law, safety, and consent
A California judge signalled that a former NPR host's voice suit against Google can go ahead (October 21). After a hearing in Santa Clara County Superior Court, Judge Shella Dean issued a tentative ruling overruling Google's demurrer on all four of David Greene's claims, which allege Google appropriated his voice and identity in an AI tool, and took the matter under submission. Greene says the male host voice in NotebookLM's AI podcasts copies his; a federal judge sent the case back to state court in July. These are allegations, and Google denies them.
Sources: Bloomberg Law, Bloomberg Law, July
The FCC said it has no plans to act on a waiver for AI political robocalls as comments closed (October 19). Club for Growth asked in August for permission to place artificial or prerecorded voice calls, including AI voices, to cellphones without prior consent before the midterms. On the last day of comments, an FCC spokesperson said: "There is no one at the FCC that has expressed any interest in moving to an order on this or any of the petitions, and there are zero plans to do so." Senate Democratic leader Chuck Schumer urged the agency to reject it, saying campaigns "could use their opponents' voice" in calls. Most reply comments opposed the waiver; the American Association of Political Consultants supported it.
Sources: AP via Al Jazeera, Communications Daily
Meta was sued over smart glasses that allegedly recorded conversations without "Hey Meta" (October 19). The proposed class action, Subia v. Meta Platforms in the Northern District of California, says Ray-Ban Meta and Oakley Meta glasses captured and saved conversations without the wake phrase, in breach of California's wiretapping law, and seeks $5,000 per violation for California buyers. Users had reported recordings and transcripts they did not ask for after a firmware update. On October 15 Meta said "This issue has been resolved" but gave no cause or count of affected users. These are allegations.
Sources: Open Class Actions, Malwarebytes
A judge dismissed or sent to arbitration the claims that Google's contact center AI eavesdropped on calls (October 22). In the Google Cloud Contact Center AI privacy litigation, Judge Rita F. Lin of the Northern District of California dismissed the claims of Home Depot callers, finding they had not shown Google intended to listen without consent; they may amend by November 12. Callers to Verizon, Uber and Discover were sent to individual arbitration under those companies' terms. No class was certified.
Source: Open Class Actions
Scammers cloned a Florida sheriff's captain's voice for fake warrant calls (October 19). The Osceola County Sheriff's Office said callers used an AI clone of Captain Steve Moser's voice, and his photo, to tell residents they had missed jury duty or had an active warrant and must pay. Sheriff Chris Blackmon said: "They cloned his voice and was using Steve's picture as well to say it was him making the phone call, and it was not." The office counts about 40 law enforcement impersonation cases this year and says it never calls to demand money.
Source: WKMG
Denmark tabled a bill that bans publishing realistic voice and face imitations without consent (October 8, missed earlier). Bill L 49, from culture minister Zenia Stampe, amends the Copyright Act to protect a person's appearance, voice and movements, and puts the burden of proving consent on whoever publishes the imitation. Protection runs until 50 years after death. Satire and parody are exempt unless they are misinformation that can cause serious danger, and platforms must take content down quickly. The government aims for it to take effect on January 1, 2027; it still has to pass the Folketing.
Also this week:
VoiceBox (VB Assets) sued Roku in Delaware federal court over seven voice assistant patents, accusing the Roku Voice Remote, Voice Remote Pro and mobile app (October 16, missed last week). Three of the patents were in the case that ended in a $40,007,600 judgment against Amazon's Alexa. Sources: VB Assets, Patent Litigator
Punjab National Bank warned customers that fraudsters may clone the voice of a family member or friend to fake an emergency, and told them to call back on a saved number (reported October 20). Source: India Business Trade
In the wild: products, enterprise, and culture
Wyndham's AI phone concierge now answers calls at more than 1,700 hotels (October 22). On the company's third-quarter earnings call, CEO Geoff Ballotti said Wyndham AI Concierge, which he called "our AI-powered voice platform", is autonomously handling calls at more than 1,700 hotels worldwide, up 17% from the previous quarter, and has answered more than 3 million calls so far this year. He said it books reservations at an average daily rate 15% above hotel book rates and has added more than 500 basis points of direct contribution for franchisees by catching calls that used to go unanswered. These are company figures.
Source: Wyndham Q3 2026 earnings call transcript (The Motley Fool)
Yelp will phone restaurants to check the wait for you (October 20). With Check for Me, a user enters party size and phone number, an AI agent calls the restaurant, and the wait time comes back by text within minutes. It will roll out "in the coming months" on eligible restaurant pages and in Yelp Assistant, in the US and on iOS only. It was part of a fall release of more than 30 features.
Sources: Yelp, TechCrunch
Google put Gemini voice into Google Maps on Apple CarPlay (October 21). During navigation, drivers can tap a Gemini icon and ask multi-step questions, change the route, for example to avoid tolls, or report traffic problems by voice. It is the first voice assistant Google Maps has offered in CarPlay. It is available now in the US, with a global rollout under way. A day earlier, Google added opt-in Personal Intelligence to Gemini in Android Auto, which lets the assistant draw on Gmail, Calendar and other Google apps while driving. Google says it will reach cars with Google built-in in the coming months.
Sources: Google, 9to5Google
GM is turning OnStar into a voice assistant, starting with the Cadillac Lyriq (October 21). Drivers say "Hey OnStar" to change the cabin temperature, open the glove box or charge port, check range and charging stops, or ask what a warning light means. When a request needs a person, it can hand the driver to an OnStar Advisor. It arrives on the 2027 Cadillac LYRIQ through an over-the-air update in the coming months, with other GM vehicles to follow. A new OnStar app in 2027 will replace the current OnStar Guardian, Buick, Cadillac, Chevrolet and GMC apps. TechCrunch reports the assistant is an in-house GM build.
Sources: GM, TechCrunch
The Diary of a CEO dubbed its Zelenskyy interview with an AI copy of his own voice (October 22). Zelenskyy spoke Ukrainian in the interview, recorded in Brussels on October 16, with two human interpreters in the room. The English track was produced with AI voice technology built from his voice, with his permission, from a human translation that interpreters and a Ukrainian editor checked. The team says it kept his pauses, breathing and Ukrainian accent. The episode was also released in more than 20 other languages. The show has not named the voice vendor.
Source: Steven Bartlett
An indie developer pulled AI voices from his game after a backlash and offered refunds (October 23). Paolo Pedercini (Molleindustria) updated his ecosatire game Future? No Thanks! so that two robot characters, an automated toilet and a hair salon translator, are now voiced by human actors instead of text-to-speech. The change followed criticism on Bluesky after he disclosed his use of generative AI. He offered to refund any buyer who "still feel[s] betrayed".
Sources: Molleindustria, Rock Paper Shotgun
A Vermont man with ALS got his voice back from 16 seconds of a 2012 TV segment (October 23). Paul Francoeur, 67, had no usable recordings from before he lost his speech. Local station WCAX found a "Made in Vermont" story it had filmed with him 14 years ago. ElevenLabs built a clone of his voice from 16 seconds of that footage, and it now replaces the synthetic voice of his eye-gaze communication system.
Source: InvestigateTV
Deployments this week: (October 19 to 23)
Synthflow AI voice agents went live on the Zoom App Marketplace for Zoom Phone and Zoom Contact Center. Batteries Plus runs them on Zoom Phone across more than 700 US stores, and its CIO says the rollout took five weeks from kick-off to pilot (October 20). Source: Synthflow (Business Wire)
Deutsche Telekom launched Voice AI Notes Mobile, which records, transcribes and summarizes business mobile calls for small and medium-sized businesses. It runs on AudioCodes VoiceAI Connect in Telekom's T Cloud and connects the mobile network to Mistral AI services (October 19). Source: AudioCodes (PR Newswire)
Procore released AI Go in open beta, a push-to-talk app that gives construction crews spoken answers from their drawings, specs, RFIs and daily logs. A Meta partnership will bring it to AI glasses with "Hey Procore" (October 21). Source: Procore (Business Wire)
Teleste built Acapela's AI text-to-speech voices into its S-ARRIVE passenger information system for trains, trams and metros, with pronunciation tuned to local station names. Teleste says it is the first to do so (October 21). Source: Teleste (PR Newswire)
Orange and Elisa are running voice services on NVIDIA hardware at the network edge. Orange uses Nemotron Speech for real-time multilingual translation for emergency and field teams, and Elisa runs an enterprise voice assistant (October 21). Sources: NVIDIA, The Fast Mode
Zoom expanded Clinical Notes, which drafts clinical documentation from in-person, video and phone visits using Zoom's own medical speech recognition. It takes speech in 10 languages and writes notes in English (October 20). Source: Zoom
SyncWords said it live-dubbed 52 Eisbären Berlin ice hockey games from German into English for a US streaming channel, using speech recognition, translation and AI voice, for AEG Sports. It showed the work at Sportel in Monaco (reported October 23). Source: Broadcast
Tellia opened Tellia Lite, a $20 per user per month voice note app for independent agronomists and growers that works offline on iOS and Android (October 19). Source: Vinetur
Papers this week
iFlytek's Spark-Audio 1.0 was trained entirely on Chinese-made chips (arXiv, October 22). The model pairs a 0.65B-parameter audio encoder with a 30B-A3B mixture-of-experts LLM, was trained on 13 million hours of audio, and covers recognition in 99 languages and 202 dialects; training ran on Huawei Ascend 910B and Hygon BW1000 hardware. The team reports performance comparable to Qwen3.5-Omni-Flash on most audio and text tasks with less audio training data (vendor figures). Source: arXiv
A speaker's own voice can trigger a backdoor in speech recognition (arXiv, October 19). Imperial College London researchers fine-tuned models so that one target speaker's natural voice activates an attacker's payload, reaching 90.2% backdoor success on Whisper-medium and 97.0% on Parakeet with false-positive rates of 1.7% and 0%, using as little as two minutes of the target's audio. When the tampered transcripts were passed to GPT-4o and Claude agents, the injected instruction drove the harmful action in up to 87.5% of cases. Sources: arXiv, Help Net Security
Full-duplex speech models break down in long conversations (arXiv, October 23). KAIST researchers tested five recent full-duplex models on the new FD-LHB benchmark, and every model eventually degraded or ran out of context, with sliding-window models falling into repetition, silence and gibberish. Two training-free fixes, keeping the first KV-cache entry and a duplex repetition penalty, reduced collapse in replays of up to an hour while running in real time. Source: arXiv
Omni models struggle with who said what in long multi-speaker video (arXiv, October 21). Omnilingual-MsASR, from Alibaba and Northeastern University (China), covers six languages, recordings up to 30 minutes and up to 20 speakers with parallel audio and video; the tcpWER of leading omni-modal LLMs frequently exceeds 100%. Source: arXiv
Real-time models answer fast but often miss the moment to respond (arXiv, October 20). The DIVE benchmark tests full-duplex interaction with continuous video across 525 annotated audio/video samples; leading real-time models reach sub-second average response latency, but their average window response rates stay at or below 53.77%. Source: arXiv
Text-only LLM correctors damage ASR transcripts that were already right (arXiv, October 21). With Gemini 2.5 Flash, correcting its own draft from text alone broke 13 to 32% of already-correct transcripts, against 1 to 3% when the corrector also had the audio; the open-weight Voxtral, which rarely edits, did not show this pattern. The authors note prompts differ between the two setups, so the comparison bounds rather than isolates the effect of audio. Source: arXiv
Deepfake detectors lose accuracy once speech goes through a real phone call (arXiv, October 20). RealComm pairs digital speech with the same utterances transmitted over real mobile calls; existing detectors degrade substantially after transmission, and combining simulated augmentation with real-call training cut the equal error rate of Teffic-Audio from 37.94% to 22.37%. Source: arXiv
Monsoon collects spontaneous speech for ASR in 50 languages (arXiv, October 19). Across seven 3,500-hour corpora, the framework's quality stages reduce semantic WER by 10 to 39% relative to a conventional protocol at equal volume, and Whisper fine-tuned on Monsoon reduces semantic WER by 20 to 44 points on conversational benchmarks; the framework, quality-control models and fine-tuned models are released. Source: arXiv
Embedding similarity disagrees with listeners on cloned dubbing voices (Information, October 20). In a benchmark of four TTS systems dubbing Chinese short dramas into Japanese, GPT-SoVITS was the only cloning system with usable intelligibility (CER 0.215, against 1.152 for F5-TTS and 4.603 for CosyVoice2). F5-TTS scored slightly higher on embedding similarity (0.729 vs 0.722), yet five native-level raters preferred GPT-SoVITS three to one. Source: MDPI
Previous edition: Voice AI News, Week 42, October 12 to 18, 2026