How to monitor brand mentions in podcasts, radio, and video with speech-to-text

Audio brand monitoring means transcribing podcasts, radio, TV, and video, then scanning the transcripts for mentions of your brand, products, and competitors. A speech-to-text API turns hours of audio into searchable text. Speech Understanding models then detect the brand names, the sentiment around each mention, and the topic of the segment, so you can alert on the mentions that matter.

Text-based social listening misses a lot. A host's offhand recommendation, a competitor's ad read, or a guest's complaint never appears in a tweet. If it's spoken, you need audio in the pipeline. This guide shows how to build that pipeline with AssemblyAI: which models to use, how to catch brand names spelled correctly, and how to turn transcripts into alerts.

Why audio brand monitoring is hard

Most media monitoring tools were built for text. Audio adds four problems:

• Proper nouns: brand and product names are often rare words. A generic model may transcribe "Notion" as "notion" or a startup name as three random words, and your keyword search misses it.
• Volume: broadcast and podcast monitoring means thousands of hours a week. Manual listening doesn't scale.
• Context: "We dropped Brand X" and "Brand X sponsors this episode" are both mentions, with opposite meaning.
• Language: global coverage means non-English and mixed-language audio, often in the same segment.

The audio brand monitoring pipeline

1. Ingest: pull episodes from RSS feeds, record broadcast streams, or collect video files.
2. Transcribe: convert speech to text with an accurate model, primed with your brand vocabulary.
3. Detect: find brand and organization mentions with Entity Detection, plus exact-match search on your keyterms.
4. Enrich: add sentiment per sentence and topic labels per segment.
5. Alert: send mentions that match your rules (negative sentiment, competitor plus your brand, a specific show) to Slack, email, or your dashboard.

Step 1: Transcribe with brand-aware speech-to-text

Start with Universal-3.5 Pro, AssemblyAI's pre-recorded flagship. It costs $0.21/hr and posts a 4.35% average WER on our published benchmark, the lowest of the systems listed. It also handles native code-switching across 18 languages, so a bilingual podcast doesn't need two passes.

The bigger win for brand monitoring is keyterms prompting. Pass up to 1,000 brand, product, competitor, and spokesperson names with keyterms_prompt (+$0.05/hr). The model then favors those spellings when it hears them. That's the difference between finding a mention and missing it. The prompting and keyterms docs cover the limits.

Need coverage beyond 18 languages? Universal-2 ($0.15/hr) supports 99+ languages.

Step 2: Detect brands, sentiment, and topics with Speech Understanding

The Speech Understanding API runs task models on top of the transcript in the same request:

Speaker labels matter more than people expect. A mention from the host carries different weight than one from a caller, and ad reads often come from a different voice.

import time
import requests

API_KEY = "YOUR_API_KEY"
HEADERS = {"authorization": API_KEY}
BRANDS = ["Acme Cloud", "Acme Pay", "RivalSoft"]

job = requests.post(
    "https://api.assemblyai.com/v2/transcript",
    headers=HEADERS,
    json={
        "audio_url": "https://example.com/episode-142.mp3",
        "speech_models": ["universal-3-5-pro"],
        "keyterms_prompt": BRANDS,
        "speaker_labels": True,
        "entity_detection": True,
        "sentiment_analysis": True,
        "iab_categories": True,
    },
).json()

while True:
    t = requests.get(f"https://api.assemblyai.com/v2/transcript/{job['id']}", headers=HEADERS).json()
    if t["status"] in ("completed", "error"):
        break
    time.sleep(5)

for s in t["sentiment_analysis_results"]:
    for brand in BRANDS:
        if brand.lower() in s["text"].lower():
            print(f"[{s['start'] // 1000}s] {brand} | {s['sentiment']} | Speaker {s['speaker']}: {s['text']}")

Sample output:

[1834s] Acme Pay | NEGATIVE | Speaker B: Honestly, Acme Pay's checkout flow lost us a week.
[2210s] RivalSoft | POSITIVE | Speaker A: RivalSoft's new dashboard is the best thing they've shipped.

Step 3: Summarize and route mentions with LLM Gateway

Raw mentions are noisy. Send each matched segment, plus a minute of surrounding transcript, to LLM Gateway. It's a single API to models from Anthropic, OpenAI, Google, and others. Ask it to classify the mention (sponsorship, recommendation, complaint, comparison) and write a one-line summary for the alert. You get structured, explainable alerts instead of keyword hits.

Monitoring live broadcasts

For radio, live TV, and livestreams, use the Streaming API with Universal-3.6 Pro Realtime ($0.45/hr). It covers 32 languages and includes keyterm prompting for up to 100 terms at no extra cost. On a 12,460-scenario benchmark published September 29, 2026, it posted a 14.4% entity error rate, compared with 18.5% for ElevenLabs Scribe v2 and 26.1% for Deepgram Nova-3. That matters when the entities are your brand names. Pipe final turns into the same detection and alerting logic you use for recorded audio. Product details are on the Realtime Speech-to-Text API page.

Monitoring non-English and mixed-language audio

Pick the model by language coverage:

• Universal-3.5 Pro (pre-recorded): 18 languages with native code-switching, so mid-sentence language changes stay in one transcript.
• Universal-3.6 Pro Realtime (live): 32 languages with automatic detection.
• Universal-2 (pre-recorded): 99+ languages for long-tail coverage.

Add the local spelling of your brand and product names to keyterms_prompt for each market.

What to build next

Once the pipeline works on one show, scale it out. Subscribe to RSS feeds for every podcast in your category. Store mentions with timestamps so your team can jump to the exact second. Then trend mention volume and sentiment by show, by week, and against competitors. We're publishing a step-by-step tutorial for a podcast brand-mention alerter soon. For the broader picture, read what is media monitoring and what is automatic speech recognition.

Frequently asked questions

How do I monitor podcasts and radio for brand mentions?

Transcribe the audio with a speech-to-text API, priming it with your brand names through keyterms prompting. Then detect mentions with entity detection and keyword matching, and add sentiment. AssemblyAI does all of this in one request with Universal-3.5 Pro and the Speech Understanding API. Live radio uses the Streaming API.

What are brand monitoring tools?

Brand monitoring tools track where and how your brand is mentioned across news, social, and broadcast media. Most track text. Audio brand monitoring adds speech-to-text so podcasts, radio, TV, and video are searchable too.

Can AI transcribe live broadcasts for media monitoring?

Yes. AssemblyAI's Universal-3.6 Pro Realtime transcribes live audio streams in 32 languages, with keyterm prompting included for brand names. Final turns can feed the same mention detection and alerting you use for recorded audio.

How do I make sure brand names are spelled correctly?

Use keyterms prompting. Pass your brand, product, competitor, and spokesperson names in keyterms_prompt: up to 1,000 terms on Universal-3.5 Pro (+$0.05/hr), or up to 100 on Universal-3.6 Pro Realtime at no extra cost.

How do media monitoring services handle non-English audio?

They route audio to models with matching language coverage. With AssemblyAI, Universal-3.5 Pro covers 18 languages with code-switching, Universal-3.6 Pro Realtime covers 32 live, and Universal-2 covers 99+.

How much does audio brand monitoring cost with AssemblyAI?

Universal-3.5 Pro is $0.21/hr. Typical add-ons are keyterms prompting (+$0.05/hr), Entity Detection (+$0.08/hr), Sentiment Analysis (+$0.02/hr), Topic Detection (+$0.15/hr), and speaker diarization (+$0.02/hr). That's $0.53/hr of audio with everything on. See the pricing page.