Fabrizio Ferrari

⚡️

[DaDaScribe @DaDaScribe]

Students watch less than a quarter of recorded lectures. That is the headline number from a crossover study at a UK Russell Group university: average viewing of captured lectures ran 23.7 percent early in the semester and 18.1 percent later.

The same study found captions and transcripts did not change how much students used the recordings. But students rated them among the most useful features available, and the functions they actually used tell you why: jumping to specific sections via the time bar, pausing, and adjusting speed. Bookmarking and built-in study notes went almost untouched.

The practical read for students: treat the transcript as an index, not a reading assignment. Search it for the concept you missed, jump to that timestamp, watch two minutes, move on. That is targeted retrieval, and it beats rewatching an hour of video to find one definition.

For educators, the lesson is different. Providing transcripts does not change behavior on its own. Teach the search-and-jump workflow explicitly in week one, because students in the study kept missing features that were one click away. The study found push reminders did not move usage either, so the fix is habit, not notification.

#HigherEd #Studytips #SpeechToText

[InfiniStrategy @InfiniStrategy]

OpenAI is merging the Chat and Work tabs in ChatGPT, and starting in December, Plus, Pro, and Business plans will share one token allowance across ChatGPT and Codex. Chat and agentic coding will draw from the same pool, with Instant chat staying unlimited.

This is a quiet but important change in how AI usage gets priced. Until now, coding agents lived in their own bucket, so heavy Codex sessions never touched your chat quota. A shared pool makes the tradeoff explicit: every agent run competes with your everyday questions. OpenAI says most users will not notice, and that may be true today. But agent workloads consume tokens at a different scale than chat, and as agentic features spread through the main interface, the shared allowance becomes the real product limit.

The interesting part is what this signals. Usage differentiation is moving from plan tiers to task types. Watch whether OpenAI introduces task-weighted pricing next, where an agent hour effectively costs more than a chat hour. For teams standardizing on ChatGPT, budget forecasting just got harder, not easier.

Sources:
https://t.co/OxCjrZEKOa
https://t.co/Z9Ejo244dq
https://t.co/2frbVRX1MT

[Fabrizio Ferrari @fabferrari71]

@SawyerMerritt @SpaceX My Grok Bot usage got back to normal this morning. Did they change their mind?!

[Fabrizio Ferrari @fabferrari71]

@distortgeekin Mine too, back to normal ;)

[Fabrizio Ferrari @fabferrari71]

@GregGaskell I got my usage halved today. It looks like I got back to normal!

[Fabrizio Ferrari @fabferrari71]

@farzyness Yeah, but at least they should have notified their users about the cut.

[Fabrizio Ferrari @fabferrari71]

@DogecoinNorway @elonmusk Very disappointing; I really didn't expect this. Now I have to revise the whole workflow, hoping I'll stay within budget. No cool.

[Fabrizio Ferrari @fabferrari71]

Great improvements in Dadascribe!

[DaDaScribe @DaDaScribe]

Overlapping speech is the hardest problem in podcast transcription, and a new generative model just moved the ceiling.

Soloist, posted on arXiv October 14, separates individual voices from a raw mix without being told how many people are talking or which voice to target. It learns who speaks when from natural acoustics, then uses reinforcement learning on real recordings, with a judge model scoring candidates on voice purity and content preservation.

The results that matter for creators: clean speech recovered from multi-speaker overlap and noisy rooms, singing vocals pulled out of music beds with pitch accuracy intact, and separated outputs assembled into continuous speaker tracks for real conversations.

We at DaDaScribe see the same failure pattern weekly, which is why our latest system update targets it directly. Multi-speaker transcriptions improved a great deal in that release. Overlapping speech remains the hard case, but it is now greatly mitigated rather than fatal: the pre-processing cleans the mix before recognition, so labels survive most crosstalk.

Try it on your messiest panel recording: https://t.co/4X0QE4zmlH

The research is open and the direction is clear. Overlap is becoming a solved problem, one paper at a time, and messy-room shows will benefit first.

#Podcasting #ContentCreators #SpeechToText

[InfiniStrategy @InfiniStrategy]

n8n published another security bulletin on October 28, and self-hosted users should treat it as urgent. The batch covers 12 advisories: three rated High, including deprecated Code Nodes that allow code execution in the main process, insecure GraphicsMagick configuration, and insufficient isolation in workflow expression evaluation. Nine more are Medium, ranging from OAuth account hijack via unauthenticated callback to SQL injection in the Vector Store node and credential headers leaking on cross-origin redirects.

Fixed versions are v1.123.89 for the v1 branch, v3.0.1 for v3, v2.45.5 for stable, and v2.46.1 for beta. Cloud instances are patched automatically. Self-hosted operators need to move now, and this is the second bulletin this month after the October 14 batch of 15 advisories.

The take: self-hosting n8n trades vendor fees for an ops obligation that is growing. Two security waves in one month, with flaws that expose credentials and allow code execution, means a self-hosted instance is now a real attack surface, not just a cost optimization. If you run n8n in production, patch on bulletin day, restrict who can create workflows, and audit exposed webhooks. If nobody owns patching, the cloud tier starts looking cheap.

Sources:
https://t.co/BfZiD8A7Fo
https://t.co/IeU1x5zDZN
https://t.co/KWUCZRYty6

[Fabrizio Ferrari @fabferrari71]

Happy Halloween!

[RoboFarm @RoboFarmGame]

Halloween on the farm: the robots are not scared of anything, but today they showed up dressed as pumpkins anyway.

One of them is so committed to the costume that it is harvesting veggies with a pumpkin face, pumpkins included. The other robots are pretending not to know it.

No spooky events, no limited-time pressure. Just a cozy farm that keeps producing while you hand out candy.

Happy Halloween. Free demo: https://t.co/frsRj0Nnyp or https://t.co/GnYAWMumgX

#RoboFarm #CozyGames #IndieGame #Halloween

[DaDaScribe @DaDaScribe]

# DaDaScribe Daily Draft - 2026-10-31
Audience: Journalists, researchers, and media professionals
Topic: News about transcription/speech-to-text tools and technology

(Draft pending - agent to research and write)
Nvidia released Nemotron 3.5 Transcribe on October 28, a free open-weight transcription model that handles up to 40 languages and translates 31 of them to and from English. The company also shipped version 2 of Granary, its open speech dataset.

The specs matter for long recordings. The model processes audio up to an hour, separates up to eight speakers, and claims benchmark wins on Hugging Face's Open ASR Leaderboard. At 30 billion parameters it runs on a single chip, from Jetson AGX Thor to Blackwell datacenter hardware.

Two limits deserve attention. Transcribe works on recorded audio only, no livestreams. And Nvidia's benchmark claims are its own until independent tests catch up.

For journalists and researchers, the interesting part is what open weights enable. Transcription that runs locally keeps sensitive source recordings off third-party servers, which matters for source protection and archive retention rules. Developers can also fine-tune on industry jargon using the Granary dataset.

The release lands in a crowded month: Mistral shipped Voxtral 3, ElevenLabs launched Realtime, and Salesforce released Launa-Omni.

#Journalism #Newsroom #SpeechToText

[InfiniStrategy @InfiniStrategy]

Studio Atelico launched its Atelico AI Engine into Early Access on October 30, and it takes a different position from most AI tooling announcements. The engine runs AI on-device instead of in the cloud, works across Unity, Unreal Engine, and Godot, and handles image generation, language models, speech, and AI-driven character behavior without touching rendering performance. It ships for Windows, macOS, iOS, and Linux on both GPU and CPU.

The team behind it built their debut game Bobium Brawlers on the engine first, then opened it up. CEO Piero Molino frames the pitch as removing trade-offs: studios no longer have to choose between intelligent gameplay, performance, and creative control.

What stands out is the artist compensation angle. Atelico trains its models only on art it owns, pays its artists, and drafted an Artist Rights Contract that lets artists define how their work can be used and share in any upside. That is a sharper answer than most of the industry has offered to the training-data question.

The take: on-device AI in game engines is the practical path for indies. No API bills, no latency, no cloud dependency, and no data leaking to third parties. If the engine holds up under real production loads, the ethical training model could become a selling point rather than a cost center.

Sources:
https://t.co/Hob9b39Kj6
https://t.co/W808AG1sVJ
https://t.co/h6ivRC1cZ4

[Fabrizio Ferrari @fabferrari71]

This is phenomenal. My son had the same idea a few weeks ago, and someone made it:

https://t.co/dHcklOkZi1

[DaDaScribe @DaDaScribe]

The microphone choice for transcription comes down to one number most buyers never check: self-noise.

Condenser mics, the default recommendation for vocals, carry self-noise of roughly 4 to 15 dBA. You hear it in quiet passages, and so does the transcription model. Dynamics measure effectively zero, because their heavier diaphragm and coil generate signal without powered electronics.

That spec gap explains the whole decision. A condenser hears everything in the room: the fan, the refrigerator two rooms away, the reflection off your desk. A dynamic hears the thing directly in front of it. Condensers also need 48-volt phantom power and extend past 20 kHz; dynamics roll off near 15 kHz and run on the cable alone.

For untreated bedrooms and home offices, the dynamic wins on transcription readiness. Reflections arrive at a lower level, the cardioid pattern rejects the rear of the room, and close-mic technique tilts the voice-to-room ratio in your favor. Put the condenser in a treated space and it earns its detail.

We compared six popular mics on transcription readiness, with specs and prices, in the Learning Center: https://t.co/x84yN5fym9

Treat the room or pick the mic that forgives it.

#MusicProduction #VoiceOver #SpeechToText

[Fabrizio Ferrari @fabferrari71]

Love it, but somewhat expensive...

[InfiniStrategy @InfiniStrategy]

Twenty years after Massive defined a generation of bass music, Native Instruments just shipped its third generation. Massive XL launched October 27 with a feature that changes how patches get made: drop in any audio file, pick a four second slice, and the built-in AI model generates four fully editable presets from it. Stem separation is included, so a sample from an existing song splits into vocals, bass, drums, and other before conversion.

This is Synplant-style audio-to-patch generation inside a serious synth architecture, with variation and timbre controls plus six auto-assigned macros per preset. Absynth's preset mutation carried over too, alongside custom wavetable creation, which users have requested since the original shipped.

The pricing tells the real story. Massive XL lists at $199 with an intro price of $99 through December 31, and a free Massive XL Player handles preset playback. That is NI's first major instrument release since the inMusic acquisition, and it reads like a company fighting to stay relevant in a market where producers expect AI assistance as table stakes.

The take: sample-to-patch tools will not replace sound design, but they compress the blank-page problem. The producers who win will treat generated patches as starting points, not endpoints. Anyone who bought into the original Massive wave should check the upgrade discount before the intro pricing lapses.

Sources:
https://t.co/EqrKh2U3qP
https://t.co/4x0TlspWse
https://t.co/s1zO4DG8sa

[Fabrizio Ferrari @fabferrari71]

Just getting started…

[Neuralink @neuralink]

Our engineers recently visited one of our speech restoration study participants for a 72-hour hackathon.

The goal: rapidly prototype as many features as possible to enhance the experience of using Neuralink to communicate with thought.

Here are some of the things we built 🧵

[Fabrizio Ferrari @fabferrari71]

Awesome!

[Alex Finn @AlexFinn]

Grok Bot just had its biggest update yet

It introduces pages and dashboards, 2 things that solve Grok's BIGGEST problem

In this video I cover how I made 5 dashboards that make Grok Bot SO much nicer to use, including one that has saved me a TON of usage:

[InfiniStrategy @InfiniStrategy]

Google now answers most branded searches with an AI Overview before anyone sees your site. Ahrefs measured 82.91% of US desktop branded keyword results showing one as of September 29, up from 13.1% a year earlier. Monthly averages climbed from 61.3% in July to 73.3% in September, based on 232.8 million crawls.

This is not ChatGPT taking traffic. It is Google's own AI layer covering the searches brands assumed they owned. Typing a company name used to guarantee your homepage first. Now the answer box sits above it, and in Ahrefs' 100-brand panel, 48 of 63 branded AI Overviews held position one.

What the AI cites matters more than what you publish. Wikipedia, YouTube, and LinkedIn led external citations in the study. That means your entity data, not your blog, shapes what buyers read about you.

The practical response is unglamorous. Audit your About page, fix stale facts, keep LinkedIn and Google Business Profile current, and search your own brand weekly to see what the answer says. Branded search is no longer owned territory. It is earned, and the audit is cheap compared to losing the click.

Sources:
https://t.co/8Fg5itFsf1
https://t.co/1ElP84HEKX
https://t.co/eKWt7GjuOT

[DaDaScribe @DaDaScribe]

Timestamps are the most underused feature in a transcript. Most people treat them as navigation. Used well, they are a quality-control system.

Three habits worth adopting.

Spot-check by timestamp, not by scrolling. Pick three random points in the transcript, jump to those moments in the audio, and compare. Ten minutes of targeted checking catches more errors than a full skim, because skimming reads what you expect while comparison hears what was said.

Timestamp the disagreement, not just the quote. When a translated or transcribed passage looks wrong, note the timecode before you fix it. A pattern of errors clustering at the same points, the fast passage, the crosstalk, the code-switch, tells you where the next file will fail too.

Use timestamps to price review time. If a one-hour file shows error clusters in the first and last ten minutes, you know the intro and the wrap-up need full review while the middle needs only spot checks. That turns an hour of proofreading into twenty minutes of targeted work.

The transcript is the deliverable, but the timestamps are the map of where it can be trusted. Clients notice the difference, even when they cannot name it.

#Translation #Localization #SpeechToText