Equibles is an MCP server that gives ChatGPT and Claude earnings-call transcripts, SEC filings and market data. On November 3, 2026, I ran Amazon's Q3 2026 call through it for a tutorial: what management promised, what analysts asked, how the tone changed, and how often a word came up over eight calls. The last one taught me the most.
A word count over two years of calls
GetTranscriptPhraseTrend counts a phrase on each of a company's last calls: whole-word, case-insensitive, per 10,000 words, split between prepared remarks and Q&A. It treats a hyphen, a space and no gap as the same spelling, so "co-pilot" finds "copilot".
For "capacity" on Amazon, it worked exactly as you'd want: 4 mentions on the February 2025 call (5.0 per 10,000 words), 21 on the October 2026 call (25.0), with 14 of those 21 in the Q&A. Analysts kept asking about it.
Where it broke
Then I asked for "Trainium", Amazon's AI chip. The last two calls showed 0 mentions, though an analyst on the October call asked about "the decision to sell Tranium externally".
That's the problem: the transcripts are speech-to-text, and they spell it "Tranium". Asking for "Tranium" returned 9 and 6 mentions on those two calls. One earlier quote reads "Trinium". Normalising hyphens and spaces doesn't help with a misheard product name, and neither does case folding.
The same transcripts misspell analysts. On the Amazon call, J.P. Morgan's Doug Anmuth appears as "Doug Anmeth". On other calls the same week I found worse. A name is just a string to a speech model; it has no idea who covers the stock.
What I'd do about it (and what the model can do now)
- For the model: when a count is zero and the topic is obviously on the call, ask for spelling variants before concluding anything. The theme extraction for the same call listed "Tranium" as a keyword, which is a decent hint.
- For the pipeline: bias the speech model with a per-company vocabulary built from the company's own filings and press releases (product names, executive names, the analysts who asked last time). Amazon's 8-K release for the quarter spells it correctly: "its custom AI accelerator, Trainium, with TorchNeuron".
- For numbers: never trust a figure from a speech transcript alone. The CEO said AWS grew 39.3%; the 8-K says 39%, rounded. Both are right, but only one is the record.
The full walkthrough, six prompts with the answers, is here:
https://equibles.com/blog/how-to-analyze-an-earnings-call-with-chatgpt-or-claude
If you work on ASR for financial audio: do you bias with a vocabulary per company, or fix entities after transcription?
Data: Amazon's Q3 2026 earnings call and 8-K, processed by Equibles.
Top comments (0)