Updated 2026-10-31 13:37 UTC Join on Telegram

Topic

media 8 signals 1 in 24h

AI Voice

Speech generation, voice agents, speech-to-text, and real-time audio models.

Latest 2026-10-31 09:17 UTC 8 source-backed 0 watch RSS JSON Feed Page JSON

Latest Signals

AI Voice feed

8 on this page 8 total
2026-10-31 09:17 UTC

Measuring voice agents: turn-taking, tool calls and expressive speech

A Hacker News thread links to a blog post from soniqo.audio titled "Measuring voice agents: turn-taking, tool calls and expressive speech." The item is listed with 1 point and 0 comments, and the provided snippet gives no further details. The headline indicates the post covers how voice agents are measured across three areas: turn-taking, tool calls and expressive speech. The source material available here does not describe specific benchmark results, numbers or findings.

Hacker News CREATOR TOOLS TECHNICAL
SOURCE-BACKED 78% Open signal Original source
2026-10-29 17:04 UTC

Parakeet v3 on a desktop CPU: speech to text at 20× real time

A post on Hacker News links to a write-up titled "Parakeet v3 on a desktop CPU: speech to text at 20× real time," hosted at talkhawk.app. The headline describes running Parakeet v3, a speech-to-text model, on a desktop CPU at 20 times real time. The Hacker News entry has 1 point and 1 comment, and the provided snippet contains no further details about the method or setup. The source does not give more specifics beyond the headline and the discussion counts.

Hacker News USEFUL NOW PRACTICAL
SOURCE-BACKED 78% Open signal Original source
2026-10-20 15:30 UTC

How Vin Bui became engineer #1 at Nowadays and started treating it like his own company

According to the Y Combinator Blog, Vin Bui found Nowadays on Work at a Startup, saw a codebase he could fix, and joined as the first engineer. The post describes his work building AI voice tools and mentoring the team. The snippet frames his path as a shift from joining a company to treating it like his own. It does not give details on funding, product metrics, or dates.

Y Combinator Blog CREATOR TOOLS GENERAL
SOURCE-BACKED 92% Open signal Original source
2026-10-19 19:43 UTC

Study Proposes Speaker-Activated Backdoors That Compromise Speech Models During Fine-Tuning

A paper on arXiv describes a threat model for speech models, which are increasingly used as interfaces to AI systems such as transcription tools and voice agents. Existing audio backdoors need an attacker to introduce triggers at inference time, when the model is actually being used. The authors study a different scenario in which the attacker compromises the model during fine-tuning and has no access to the victim's inference-time inputs. They introduce the Speaker-Activated Backdoor (SAB) setting, which the source snippet cuts off before fully explaining its mechanics.

arXiv RESEARCH TECHNICAL
SOURCE-BACKED 95% Open signal Original source
2026-10-19 11:59 UTC

Heartbeat Thinking proposes speculative reasoning and tool execution during speech in voice agents

A new arXiv paper titled Heartbeat Thinking examines tool-using realtime voice agents, which often make users wait after they finish speaking even when actionable information arrived earlier in the request. The snippet says existing speculative approaches start work during speech, but chunk-bounded or interrupt-on-update generation can cut off a tool call before its arguments are complete. This ties computation to speech updates and limits how much useful work can be done in advance. The authors introduce Heartbeat Thinking as an approach to this problem, though the provided text is truncated and does not give its full method or results.

arXiv RESEARCH TECHNICAL
SOURCE-BACKED 95% Open signal Original source
2026-10-19 00:10 UTC

Study Finds AI Voice Agents Can Make Human Service Workers Less Willing to Persist With Leads

An arXiv paper examines what happens when firms use AI voice agents in high-volume contact centers to handle low-intent leads while routing high-intent leads to human workers. Drawing on sequential search theory, the authors argue that this task allocation enriches the opportunities facing human workers but also raises their reservation value for the lead in hand. As a result, the authors posit that human workers may become less willing to persist with the leads they hold. The snippet is cut off before it describes the study's methods or results, so the specific findings are not available here.

arXiv MONEY GENERAL
SOURCE-BACKED 95% Open signal Original source
2026-10-05 16:02 UTC

Study Examines Accent Gap in Speech-Driven Financial Voice Assistants

A paper posted on arXiv examines how AI voice assistants perform with British accents in financial settings. According to the source, these systems use Automatic Speech Recognition (ASR) combined with LLM-based reasoning, yet they struggle with regional British accents, including Scottish, Irish, and Welsh. The source attributes this to most ASR models being trained predominantly on American English voice data. It also says recognition errors can carry through to the LLM stage, corrupting tool-call arguments and leading to wrong or missing responses.

arXiv TECHNICAL
SOURCE-BACKED 95% Open signal Original source
2026-09-30 02:18 UTC

Study Examines How AI Voice Interviews Change What Employees Disclose in Organizational Listening

A field study inside a global management consulting firm examines how employees' disclosures change when they are asked about the same themes through a conversational AI voice interview. The firm's process pairs a pre-survey with an adaptive AI voice interview within one session. The research focuses on situations where disclosure carries hierarchical risk. The provided excerpt is cut off before it reports its main results, so those findings are not included here.

arXiv TECHNICAL
SOURCE-BACKED 95% Open signal Original source

About this Topic

Stable editorial lens

Speech generation, voice agents, speech-to-text, and real-time audio models.