Why ASR hears “Road” when the video says Rhode
Social listening is noisy on video because most tools match brands on ASR transcripts. General speech models are not trained on beauty names, so Rhode becomes Road and silent pack-shots vanish.
Social listening data is noisy on video because most tools match your brand on an ASR transcript, and general speech models are not trained on beauty or luxury names. Rhode becomes Road. Refy collapses into a rotating set of tokens. If the creator never says the name — silent ASMR, a pack-shot held to camera — the transcript has nothing to match. The dashboard still reports coverage. The index is already wrong. The fix is not a better prompt on the same audio. Index on what is on screen: visual SKU matching first, then use speech only as supporting context for claims, pronunciation, or codes.
Why social listening data is so noisy on video
General ASR is trained on news, calls, and English at large. It is not trained on beauty, fashion, or niche luxury lexicons. Once you brand-match on that transcript, every downstream number is contaminated. Share of voice includes videos that said “road.” Mentions of the actual brand disappear. The customer never sees the error because the UI still says “we found your brand in video.”
What we see when transcript matching is the primary key
Typical ASR collisions from Noodle4 R&D on beauty lexicons| On screen / said | Typical ASR match | What should be recorded |
|---|
| Rhode, spoken | Road | Rhode |
| Refy, spoken | Gibberish or a neighbour token | Refy |
| Silent ASMR pack-shot | (empty transcript) | SKU from pack geometry |
| Haul, name never said | Unrelated words in the room | SKU if the product is in frame |
The post-mortem: we stopped trusting the ears
Fine-tuning the ASR on a brand list helped a little and then failed the next launch. New SKUs, seasonal names, and creator slang do not wait for a retraining cycle. Phonetic closeness is also not brand closeness: “road” and “Rhode” are close in sound and infinitely far apart in a beauty category.
Speech is still useful as a supporting signal — a claim, a pronunciation, a discount code. It is a terrible primary key.
Visual-first SKU matching
The pipeline we kept treats vision as the index and audio as context. We match pack-shot geometry and on-screen product against the brand’s actual SKUs, then attach transcript fragments only when they corroborate what was in frame. If the audio says Road and the bronzer on the sink is Rhode, the record is Rhode.
How we know this
These notes come from building Noodle4’s review and listening pipelines, not from a published academic sample. Cost curves, intent splits, and ASR collisions are from internal R&D and client-shaped tests. They will not match every category. We include them because the failure mode is structural — not because one campaign is universal.
Alex Gray is the founder of Noodle4 and a former technical AI product owner. He works with agencies to apply multimodal AI to creator screening, draft approval, and post-publication monitoring.
Frequently asked questions
Why is my social listening data so noisy on video?
Most stacks match brands on ASR transcripts. General speech models are not trained on beauty or luxury names, so Rhode becomes Road and niche brands vanish or collide. The noise is in the index, not in the market.
How do you track a brand in video if the creator never says the name?
Noodle4 indexes on visual SKU matching and uses the transcript as supporting context, not as the primary key. Silent pack-shots and ASMR still resolve to the product on screen.