How to get earnings call transcripts as data

The short way

A call is keyed by company and fiscal quarter, so ask the vendor which quarters it holds before asking for one. With Financial Modeling Prep on a tier that carries transcripts:

import json, urllib.request

KEY = "YOUR_API_KEY"
BASE = "https://financialmodelingprep.com/stable"

def get(path):
    with urllib.request.urlopen(f"{BASE}/{path}&apikey={KEY}") as r:
        return json.loads(r.read())

# which fiscal year and quarter pairs exist for this symbol, and on what dates
dates = get("earning-call-transcript-dates?symbol=AAPL")
print(json.dumps(dates[:3], indent=2))

# then one call, by the fiscal pair the list returned — not by calendar quarter
call = get("earning-call-transcript?symbol=AAPL&year=2025&quarter=3")
print(json.dumps(call, indent=2)[:1500])

Two endpoints sit beside those and are worth knowing before you build a loop: earnings-transcript-list returns every symbol that has transcripts at all, which is how you find out whether your universe is covered before paying for the tier, and earning-call-transcript-latest returns what was added most recently, which is cheaper to poll than walking every company. Store the raw response with the date you fetched it, because a transcript you fetched today can be corrected tomorrow.

What the options are

Beside the fundamentals. Financial Modeling Prep puts transcripts on its Ultimate tier, $149 a month billed annually, together with global coverage, ETF holdings and bulk delivery — so if the transcript is one input in a model built on statements, it is the same key and the same integration. ROIC.ai claims seventeen-plus years of call transcripts beside forty years of statements, updated daily; its API history is capped by plan, at two years free, five on the $29 Individual plan and all of it on the $89 Professional one. On both, the published plans are for your own use: display or redistribution is the quote-only Enterprise tier.

Built for the call itself. Quartr API is the specialist: live audio, a live transcript streamed as JSON Lines with per-word timings, a raw transcript after the call, an edited one with speakers named, and the slide deck and report from the same event, across 16,000-plus companies. Delivery is REST, signed webhooks or a Snowflake share, and its delivery times are written into the contract. Pricing is quote-only, and internal use and redistribution are two different sets of terms rather than two sizes of one plan.

Retrieved, not stored. Bigdata.com searches premium news, filings and transcripts to ground a language model and bills per token. Its terms grant a grounding right for inference time only: you may keep what your model produced, not the content it read. That is the right shape for an assistant answering a question about last night's call, and the wrong one for building a corpus.

To read rather than to parse. Several research terminals display transcripts without selling them as data. Fiscal.ai serves them through its API and MCP server too, but its transcript history there starts with each company's latest period as of 7 July 2026, which makes it a feed going forward rather than an archive. The rest of the shelf is in fundamentals and research.

Where this breaks

There is no public original to fall back on. A 10-K has one canonical copy on EDGAR, and every vendor's version can be checked against it. An earnings call has none: the company owns the recording, each transcription is that vendor's own work, and two vendors' texts of the same call differ in punctuation, in what they mark as inaudible and in where they break paragraphs. Switch vendors partway through a series and the text changes for reasons that have nothing to do with what anyone said. Pick one source per series, and keep the vendor in the key.

The licence is the product. Quartr's 2026 internal-use terms let work derived from its data leave your firm only if it is a meaningful transformation, no continuous verbatim excerpt from one document exceeds 250 words, and verbatim material is no more than 20% of the total. The same terms prohibit training, fine-tuning or evaluating any AI model on the data, and systematically caching it to build a dataset. Redistribution is a separate contract that names the one product the data may appear in. Financial Modeling Prep and ROIC.ai keep display and commercial use on Enterprise. A side project that shows transcripts on a public page is a redistribution licence before it is an engineering problem.

Speakers arrive later, and some never do. Quartr publishes transcripts in two steps: a raw transcript with paragraph breaks shortly after the call, and an edited one once speaker identification is done, with names, roles and companies. During earnings season some events are attributed sooner than others according to client interest. Name, role and company can all be null when a speaker is not clearly identified or not on the official roster, and older transcripts are never retrofitted with speaker data. Anything that separates management's prepared remarks from the analysts' questions — the most common thing people want from a transcript — depends on a field that may not be there.

The first version is not the last. The raw transcript is a machine pass; the edited one replaces it hours later, and Quartr sends webhooks for a transcript being created, updated and deleted. A pipeline that ingests the first version and never listens again keeps the transcription errors and the missing speakers forever. Store versions, and treat an update event as a replacement rather than a new document.

Fiscal quarters are not calendar quarters. Transcripts are indexed by the company's fiscal year and quarter. Apple's fiscal year ends in late September, so its "Q3 2025" call is about the quarter that ended in June. Ask for calendar Q3 across a universe and you get a mix of quarters that ended in different months. Resolve each company through its dates endpoint, and join on the call date or the period end rather than on the label.

Coverage thins exactly where the information is. Large caps are transcribed by everyone; small caps, recent listings and companies that hold calls only some quarters are where vendors differ. Quartr transcribes only events conducted in English, a large hole in a product that covers 65 markets. Check your universe against the vendor's symbol list before you commit, and count the calls missing per quarter rather than the companies covered.

Fast is a percentile, not a promise. Quartr's contractual SLA is a transcript for 95% of events within 45 minutes of the call ending and edited transcripts within a couple of hours. The other 5% is not a rounding error if your event is in it. A vendor that updates daily can lag a full day. If the use is reacting to the call while it happens, only a live transcript does that, and it is a different dataset at a different price.

If you outgrow this

When you need the call as it happens, that is Quartr's live dataset — audio and a streamed transcript, transcribed within five seconds of event start for 90% of events under its SLA — from the same vendor and the same API as the archive.

When what you want is the numbers said on the call, transcripts are the slow way to get them. Daloopa extracts KPIs and guidance from filings, decks and selected transcripts into a model, each figure linked to its source page.

When you want the date of the next call rather than the text of the last one, see How to get an earnings calendar as data.