Nov 18, 2026
Audio WG Bi-weekly
Invited Audio AI Work Group
Summary
Working group restart reviewed past automatic speech recognition work and planned modular publications for fiscal year 2027.
Participant Introductions and Restarts
New and returning members introduced their backgrounds in edge devices, artificial intelligence, and speech technologies. The session marked a formal restart of the working group activities.
Review of Previous Deliverables
Members discussed prior automatic speech recognition landscape mapping and recovered missing project files. The massive 50 page white paper will be split into modular publications.
Prioritizing Future Topic Deliverables
Participants explained why automatic speech recognition serves as a critical front end for edge large language models. The group agreed to stack rank subtopics for fiscal year 2027 delivery.
Decisions
Aligned
- Working group charter and approach refresh The participants agreed to update the wiki with a refreshed approach and charter prior to the next meeting.
- ASR topic completion before subsequent priorities The group agreed to finish the current ASR paper deliverable before pursuing other stacked-ranked topics.
- Comprehensive survey publication as a live wiki draft The participants agreed to publish the large survey document on the wiki as an editable, live draft article.
Next steps
[Pete Bernard] Publish the full white paper: as an editable wiki article with work-in-progress caveats so participants can review the complete scope.
[Pete Bernard] Publish the draft document on the wiki: as a live working source for the audio AI working group.
[Tomer Badug] Speak with Jonathan offline: regarding potential participation and contributions to the ASR working group discussions.
Details
Participant Introductions and Working Group Restart: Tomer Badug re-engaged with the working group as an independent participant specializing in edge device audio AI after previously working with Civa (00:00:12). Yuval Ramot joined as Tomer Badug’s replacement from Civa, bringing machine learning and artificial intelligence experience (00:02:21). Abigail Oppong introduced themselves as a researcher focusing on natural language processing speech technologies for low-resource languages and upcoming atypical speech doctoral studies. Gabriele Gallo Stampino joined from the Italian startup Neuronova, bringing a software background to edge product development (00:03:51). Pete Bernard opened the session as the head of the AGI Foundation to restart the working group’s activities. Venkat Rayudu also joined the meeting late during the introductions (00:05:05).
Review of Previous ASR Application Work and Deliverables: Tomer Badug inquired about the previous progress on automatic speech recognition applications, noting that Abigail Oppong had been leading the automatic speech recognition judging efforts. Abigail Oppong recalled that the group previously worked on mapping the automatic speech recognition landscape and market needs, though records were stored on Tomer Badug’s former Civa computer (00:06:07). Tomer Badug requested that relevant files and past emails be reshared to restore access to previous assignments and research data (00:07:27) (00:11:54).
Transition of Comprehensive White Paper into Modular Publications: Tomer Badug explained that the original audio artificial intelligence edge survey white paper grew to roughly 50 or 60 pages, prompting the group to split it into separate topic-specific white papers, blog posts, and deliverables such as pre-trained models for the foundation models website model hub (00:08:40). Pete Bernard confirmed that the comprehensive document was never officially published due to its massive size (00:09:59). Tomer Badug and Abigail Oppong reshared the draft documents directly within the meeting chat for participants like Yuval Ramot, Gabriele Gallo Stampino, and Pete Bernard to review (00:11:12) (00:22:09).
Context and Rationale for Starting with ASR: Gabriele Gallo Stampino asked why the group chose automatic speech recognition as the initial subtopic to focus on. Tomer Badug and John Maruni from EmbedYOU had previously identified automatic speech recognition as a critical front end for communicating with large language models and small language models on smart, resource-constrained edge devices (00:14:07). Tomer Badug noted that while other topics were more mature, deploying automatic speech recognition on edge devices still involved complex challenges that the working group aims to make accessible for developers (00:17:10). Gabriele Gallo Stampino contrasted this with Neuronova’s focus on neuromorphic computing, wake-word detection, and ultra-low power consumption for wearable devices (00:18:44).
Prioritizing Future Topics and Planning Deliverables for Fiscal Year 2027: Pete Bernard suggested resharing the subtopics from the original white paper to stack-rank them and prioritize future efforts heading into fiscal year 2027 (00:18:44). Pete Bernard emphasized the importance of maintaining a regular cadence of deliverables, including white papers, educational content, and reference materials to increase the group’s impact (00:19:47). Tomer Badug reminded the participants about a previously published wake-word detection model produced by the working group (00:21:00).
Publishing the Comprehensive Draft on the Wiki: Pete Bernard volunteered to take the action item to publish the comprehensive 50-page survey document as a live, editable wiki article with appropriate work-in-progress caveats so all working group members can access and edit it. Yuval Ramot noted that they had only reviewed about half of the document so far (00:23:02). Pete Bernard explained that the baseline charter of all foundation working groups is educational outreach to keep members informed about space developments, with optional activities like live streams, contests, and benchmarks depending on group interest (00:23:55). Pete Bernard committed to updating the wiki and charter before the next meeting to reflect the refreshed approach (00:25:02).