Continue Reading

Transcribing 2.5M Hours Of TV News: A First Look At Chirp + CLD2 Applied To A Chinese News Broadcast

Earlier this month in collaboration with the Internet Archive's TV News Archive, we completed the machine transcription of its complete…

Continue Reading

Transcribing 2.5M Hours Of TV News: How Television News Across The World Relies More On Subtitles Than Dubbing For Multilingual Speech

When American televisions news channels broadcast a clip of someone speaking a language other than English, they typically begin with…

Continue Reading

Transcribing 2.5M Hours Of TV News: Chirp + Language Detection – English In A Persian Broadcast

As we continue our experiments in applying language detection to our multilingual Chirp speech recognition results, we keep finding fascinating…

Continue Reading

Transcribing 2.5M Hours Of TV News: Chirp + Language Detection – A Half-English Half-Portuguese Broadcast

As we continue our experiments in applying language detection to our multilingual Chirp speech recognition results, we continue to find…

Continue Reading

Transcribing 2.5M Hours Of TV News: First Experiments With Applying Language Detection To Chirp's Multilingual Speech Transcription

Last September we examined how Google's new Universal Speech Model called Chirp was the first automated speech transcription system we…

Continue Reading

Some Excerpts Of Biden's State Of The Union Address On Russian Television News

How was President Biden's State of the Union Address (SOTU) covered on Russian television news? Below are a few clips…

Continue Reading

Transcribing 2.5M Hours Of TV News: 4.25M Global Broadcasts Processed Using GCP's Chirp LSM

Two weeks ago we unveiled the first glimpse of our massive collaboration with the Internet Archive's Television News Archive to…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Using Prompt Recommendations To Expand Our Textual Descriptions

As we continue our experiments in video description using Google's Gemini 1.5 Pro, let's ask Gemini for help in crafting…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Converting Videos To Text For Universal RAG & Summarization

Over the past week we've been exploring Google's Gemini 1.5 Pro model's native video support through a series of experiments,…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Returning To Second-By-Second Visual Descriptions Through Video & Surrogates

As continue to examine Gemini's video capabilities, let's return to our work on second-by-second visual descriptions using a series of…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Boosting Gemini To 2.5 Hours Of Video + Summarization

Yesterday we explored boosting Gemini's video capabilities from its native one hour to 2.5 hours by overriding its sampling process…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Overriding Gemini's Sampling To Extend Its Context Window To 2.5 Hours

Google's new Gemini 1.5 Pro LMM model accepts videos up to one hour in length and internally samples them into…

Continue Reading

More Experiments With Gemini's Video Capabilities: Pushing Beyond The 1 Hour Limit

Stay tuned tomorrow for a first look at pushing Gemini beyond its default one hour input limitation for videos by…

Continue Reading

#MeToo Has Largely Faded Away On Television News

The term #MeToo burst into the media in late 2017 , but largely faded from CNN and MSNBC by late…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Using Image Surrogates To Summarize A Russian TV News Clip Into Stories

Yesterday we demonstrated how Google's Gemini 1.5 Pro's video analysis capabilities make it possible to take a 6 minute clip…

Continue Reading

Why Gemini 1.5 Pro Describes President Biden As "Confused" & "Disoriented" And What It Tells Us About Training Data

One of the more fascinating findings from our experiments with Gemini 1.5 Pro's video capabilities is how strongly it associates…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Visually Summarizing A Russian TV News Clip Into Stories

As we continue to explore Gemini 1.5 Pro's video analysis capabilities, we've examined a number of different prompt approaches to…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Visually Summarizing A 30-Minute Evening News Broadcast In One Prompt

As we continue to explore Google's Gemini 1.5 Pro model's native video capabilities, what if we scale our analysis from…

Continue Reading

Gemini 1.5 Pro's 1 Million Token Model: Can Prompt Engineering Improve Its "Needle In A Haystack" Performance?

Yesterday we were surprised to discover sharply reduced performance from Gemini 1.5 Pro on our RW-NIAH (Real World Needle In…

Continue Reading

Gemini 1.5 Pro's 1 Million Token Model: How Its "Needle In A Haystack" Performance On One Broadcast Collapsed to 0% Accuracy Over The Last 2 Weeks

Two weeks ago we examined Gemini 1.5 Pro's NIAH (Needle In A Haystack) performance, finding that when asked to identify…

Continue Reading

Early Findings With Gender, Racial & Cultural Bias In Generative AI Embeddings, Models & Guardrails

As we continue to explore how various generative AI offerings from different companies perform across global news coverage, one area…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: Summarizing A Video Second-By-Second With Keywords & Descriptions

As we continue our experiments with having Google's Gemini 1.5 Pro model visually analyze a 6-minute video clip from Russian…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: The Instability Of Summarizing Video

Yesterday we explored Google's Gemini 1.5 Pro model's video analysis capabilities by asking it to analyze a 6-minute clip of…

Continue Reading

LMMs & Google's Gemini 1.5 Pro Watching Television News: First Results From Watching A Russian TV News Clip

One of the most powerful capabilities of Google's new Large Multimodal Model (LMM) Gemini 1.5 Pro model is its ability…