Media monitoring explained: how AI media tracking and news monitoring work

Somebody just said your company's name on a morning radio show in Chicago. A podcast host spent four minutes on your pricing change. A regional news segment ran a clip of your CEO with the wrong number on screen.

None of that is text. None of it shows up in a keyword alert. And that's the gap most monitoring setups still have.

What is media monitoring?

Media monitoring is the practice of tracking, capturing, and analyzing mentions of a brand, person, product, or topic across news, broadcast, podcast, social, and print channels. Modern systems do it continuously rather than in a weekly clipping report, and the useful ones treat audio and video as first-class sources rather than as an afterthought.

The mechanical problem is simple to state. Text is already machine-readable. Spoken audio isn't, until you transcribe it. So every serious monitoring pipeline is really a speech-to-text pipeline with analysis bolted on top.

Media monitoring vs. media intelligence

Monitoring tells you a mention happened. Media intelligence tells you what it means: who said it, in what tone, on what topic, next to which competitor, and whether the volume is trending in a direction that should worry you.

The difference isn't philosophical. It's a data structure question. Monitoring produces a list of hits. Media intelligence produces structured records with entities, topics, sentiment, and speakers attached, which is the only form you can aggregate, chart, or route into a workflow. If you're building the second thing, you need the enrichment layer described further down, not just a transcript.

How does automated news monitoring work?

Automated news monitoring works in four stages: ingest the source, convert speech to text, enrich the text with structured metadata, then match and alert against your rules. Everything else is plumbing around those four steps.

In practice:

• Ingest. Pull broadcast streams, podcast RSS feeds, video files, and press-conference recordings into one queue. Live sources open a socket; archives get submitted as files.
• Transcribe. Every audio second becomes timestamped, speaker-attributed text. This is where accuracy either saves you or sinks you, because everything downstream reads the transcript, not the audio.
• Enrich. Extract people, companies, products, and locations. Classify the topic. Score the sentiment. Flag anything unsafe.
• Match and alert. Run your watchlist against the enriched records and push what matters to Slack, email, or a dashboard, with a timestamp that jumps straight to the moment in the audio.

The reason older news monitoring products feel thin is that most of them stopped at stage one and treated stages two through four as someone else's problem.

The media scanning pipeline: from raw audio to structured signals

Media scanning is the enrichment half of the pipeline: taking a wall of transcribed text and turning it into fields you can filter, count, and alert on. Five components do most of the work, and on AssemblyAI they run through one Speech Understanding request rather than five separate services.

Two practical notes. Entity detection is what makes disambiguation possible at scale, because "Apollo" as a company and "Apollo" as a mission are different rows once the entity type is attached. And sentiment scored per sentence is far more useful than sentiment scored per file, since a five-minute segment usually contains both.

For pre-recorded work, Universal-3.5 Pro is the model to run: $0.21 per hour, and it's the default async model on all accounts as of August 7, 2026. Standard diarization adds $0.02 per hour on top, which is the cheapest useful upgrade in the whole stack for panel shows and interviews.

Live broadcast monitoring: media tracking as it airs

Radio, TV, and live-streamed press events are genuinely real-time problems, and batch transcription is the wrong tool for them. If a competitor announces a recall at 9:14am, finding out at 11am is a different business outcome than finding out at 9:15am.

For live sources, streaming transcription runs over a WebSocket connection and bills on session duration rather than audio duration. Universal-3.5 Pro Realtime costs $0.45 per hour base, and two of its options exist almost precisely for broadcast conditions:

• voice_focus isolates the primary speaker and suppresses background speech and noise. Broadcast audio is rarely clean — crowd noise behind a field reporter, music beds under a radio segment, a room tone that swallows a quiet panelist. Use near-field for phone-in and headset audio, far-field for rooms and press pools. It's a $0.10 per hour add-on.
• Diarization with revision labels speakers live, then re-clusters and sends a single correction within about half a second of the stream ending. On a five-person political panel where everyone talks over everyone, that correction pass is the difference between an attributable quote and an unusable one. It handles up to 10 speakers and adds $0.12 per hour. The broader mechanics are covered in our guide to speaker diarization.

The model also runs in three modes rather than exposing low-level flags: min_latency, balanced, and max_accuracy. For noisy far-field broadcast feeds, max_accuracy is usually the right default, since a monitoring pipeline cares more about getting the name right than about shaving 80 milliseconds.

Monitoring media in more than one language

International monitoring breaks most pipelines in a specific way: coverage is fine until a speaker switches language mid-sentence, and then the transcript degrades exactly where the interesting content is.

Universal-3.5 Pro handles 18 languages with native code-switching out of the box — English, Spanish, French, German, Italian, Portuguese, Arabic, Danish, Dutch, Finnish, Hebrew, Hindi, Japanese, Mandarin, Norwegian, Swedish, Turkish, and Vietnamese — with no configuration and no second pass. Here's how that lands on a five-pair code-switching benchmark, normalized WER, lower is better:

Full methodology and dataset details live on our public benchmarks page.

Need wider coverage than 18 languages? Universal-2 handles 99+ languages at $0.15 per hour and is the documented fallback rather than a deprecated model. For pre-recorded media, translation into 100+ target languages means a Portuguese talk-radio segment and a Japanese business broadcast can land in an English-language dashboard side by side. Automatic language detection covers 130+ detectable languages and needs roughly 15 seconds of audio to commit, which matters when you're scanning short clips.

What does it cost to monitor audio at scale?

Cheaper than most teams assume, and the arithmetic is easy because billing is per second with no minimums.

A 1,000-hour archive backfill on Universal-3.5 Pro runs $210. Add standard diarization and it's $230. A single live radio stream monitored 24/7 for a month on Universal-3.5 Pro Realtime is roughly 730 hours of session time at $0.45 per hour. The cost driver in real deployments is almost never the model — it's how many live streams you keep open around the clock versus how much you batch overnight. Current rates for every model and add-on are on the pricing page.

One deployment note for regulated or EU-based monitoring: there's an EU data residency endpoint at the same price as US, so audio and transcripts stay in the EU.

Who builds media intelligence on this infrastructure

AlphaSense, the market intelligence platform, and Ipsos, the global research and polling firm, both build on AssemblyAI. So does a long list of teams turning recorded conversation into searchable, structured data across research, media, and analytics — more of them are on our customers page.

What those companies have in common isn't a feature. It's that they treat transcription as infrastructure rather than as a product feature, which means they care about the boring things: per-second billing, a stable API reference, and models that improve without forcing a migration.

The part most teams get backwards

Most monitoring projects start with the alerting rules — the watchlist, the Boolean logic, the routing. That's the fun part, and it's the part that fails last.

What actually determines whether the system earns its keep is the accuracy of the transcript underneath it, because a missed brand name is a missed alert and no amount of clever rule-writing recovers it. Build the ingest and transcription layer first, get the entity extraction right, and the alerting logic becomes almost trivial. Teams that do it in the other order spend six months tuning rules against text that was wrong to begin with. If you want the wider context on what structured audio data unlocks, our primer on voice intelligence covers it.

Frequently asked questions

How does media monitoring handle podcasts and radio?

Podcasts and radio are transcribed like any other audio source, then enriched with entity detection, topic labels, and sentiment so mentions become searchable records. Podcasts are typically handled as batch jobs from an RSS feed, since episodes drop as complete files. Live radio runs over a streaming connection instead, so mentions surface while the segment is still on air.

Can media monitoring track non-English news coverage?

Yes. Universal-3.5 Pro covers 18 languages with native code-switching, and Universal-2 extends coverage to 99+ languages for wider international monitoring. For pre-recorded media, translation into 100+ target languages lets non-English coverage land in the same dashboard as domestic coverage.

What's the difference between real-time and batch media monitoring?

Real-time monitoring streams live audio over a WebSocket and surfaces mentions within seconds, which is what you need for broadcast, radio, and live press events. Batch monitoring submits complete files and is the right choice for podcasts, archives, and anything where a few minutes of delay costs nothing. Most production pipelines run both, with streaming reserved for the handful of sources where speed changes the response.

How much does automated media monitoring cost per hour of audio?

Transcription for pre-recorded audio is $0.21 per hour on Universal-3.5 Pro and $0.15 per hour on Universal-2, billed per second with no minimums. Live streaming is $0.45 per hour base, billed on how long the connection stays open rather than on audio length. Speaker diarization and voice isolation are per-hour add-ons on top of those rates.

How accurate is speech-to-text on noisy broadcast audio?

Accuracy on broadcast audio depends heavily on whether you use voice isolation, which suppresses background noise and competing speech before the model commits to a transcript. On streaming, setting far-field voice focus and the max-accuracy mode is the standard configuration for rooms, press pools, and field reporting. Published word error rates across standard datasets are on the AssemblyAI benchmarks page.

Do I need separate tools for media scanning, sentiment, and transcription?

No. Transcription, entity detection, topic detection, sentiment analysis, and content moderation can run in a single API request rather than as a chain of separate vendors. Consolidating them matters operationally, because every extra service in the chain is another set of credentials, another latency budget, and another place where timestamps drift out of alignment.

‍