{"slug": "counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the", "title": "Counting words in earnings call transcripts: the speech-recognition spellings that break the counts", "summary": "An analysis of 5,305 S&P 500 earnings call transcripts found that speech-recognition misspellings, not the counting itself, were the main obstacle to tracking terms such as \"agentic\" and \"tariff refund,\" according to research published by Equibles. The author documented variants including \"agentec,\" \"agentech,\" \"a gentic\" and \"a genetic\" for \"agentic,\" and \"IEPA\" for \"IEEPA,\" the law behind the tariffs the Supreme Court struck down in February. An overlap check on July-to-September 2025 calls from two transcription pipelines showed similar shares — 11.7% versus 12.0% of calls using \"agentic\" and 61.8% versus 60.8% mentioning AI — which the author said justified trusting the series across the pipeline change.", "body_md": "I count words across earnings call transcripts (how often S&P 500 companies say “agentic”, “tariff refund” and so on), and most of the work turned out to be speech-recognition spellings rather than the counting itself. A few that bit me, in case they help anyone doing text analysis on ASR output of business speech:\n\n**Words that arrive split or respelled.** “Agentic” comes through as “agentec”, “agentech”, “a gentic” and “a genetic”. The first three are safe to count. “A genetic” isn’t, because it also catches real genetics talk on healthcare calls, so I only count it when the next word is AI, commerce, workflows or a similar term.\n\n**Acronyms.** “IEEPA”, the law behind the tariffs the Supreme Court struck down in February, shows up as “IEPA” and other variants, so a literal match undercounts exactly the quarters where it matters most.\n\n**New words with no settled spelling.** “Neocloud” arrives as “neocloud”, “neo cloud” and “neo-cloud”, sometimes all three on the same call, so the match has to allow a space or hyphen.\n\n**Names that are also people.** “Claude” needs an AI term within about 90 characters, because executives and board members are named Claude. FedEx’s interim CFO is Claude Russ, so without that rule FedEx would look like it discusses Anthropic’s model on its recent calls.\n\n**Product names on their owner’s calls.** “Gemini”, “Copilot” and “Llama” are counted only on other companies’ calls. On Alphabet’s, Microsoft’s and Meta’s own calls they’re product updates, not a sign that other companies are adopting them.\n\n**A change in the transcription process.** Most of my transcripts before 2026 and after come from different transcription processes, which is a quiet way to manufacture a trend. Where both were common in the same quarter (July to September 2025) they gave similar shares: 11.7% and 12.0% of calls used “agentic”, and 61.8% and 60.8% mentioned AI. Close enough that I trust the series, but I’d want that kind of overlap check before trusting any word trend across a pipeline change.\n\nThe write-up with the full counting rules: [Agentic AI Has Overtaken Generative AI on Earnings Calls - Equibles](https://equibles.com/research/agentic-ai-has-overtaken-generative-ai-on-earnings-calls-evidence-from-5-305-s-p-500-calls)\n\nHow do others handle domain vocabulary in ASR output? I’ve been using hand-written variant lists, which works for a handful of terms but doesn’t scale. I’ve wondered whether biasing the decoder with a domain word list (an initial prompt, or hotword boosting) beats correcting after the fact, and I’d be curious what has worked for people.", "url": "https://wpnews.pro/news/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the", "canonical_source": "https://discuss.huggingface.co/t/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-that-break-the-counts/191048#post_1", "published_at": "2026-10-27 00:01:12+00:00", "updated_at": "2026-10-27 00:13:22.859299+00:00", "lang": "en", "topics": ["natural-language-processing", "ai-research"], "entities": ["Equibles", "Anthropic", "Claude", "FedEx", "Claude Russ", "Alphabet", "Microsoft", "Meta"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the", "markdown": "https://wpnews.pro/news/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the.md", "text": "https://wpnews.pro/news/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the.txt", "jsonld": "https://wpnews.pro/news/counting-words-in-earnings-call-transcripts-the-speech-recognition-spellings-the.jsonld"}}