Hedge fund analysts paying USD 1,000 to USD 1,500 per expert call are getting transcripts with a 5-12% word error rate (WER), and the errors are not distributed evenly. They concentrate in the tokens that drive the investment decision: numbers, tickers, company names, and technical acronyms.

That is the structural problem with using general-purpose automatic speech recognition (ASR) on finance content. Vendor benchmarks measure something the buyer does not actually care about.

What the WER number actually represents

WER is the share of words in a reference transcript that the ASR system gets wrong, through substitution, deletion, or insertion. The benchmarks vendors quote (LibriSpeech, VoxPopuli, and similar) are built from audiobook readings and parliamentary proceedings. Clean audio. Native speakers. General vocabulary.

Deepgram's published benchmarks, AssemblyAI's Universal-2 announcement, and the Whisper paper from OpenAI all report results against this kind of corpus. The numbers are honest, in the sense that the benchmark is what it is. They are also close to irrelevant for an analyst transcribing a 60-minute expert call with a former TSMC fab engineer who switches between English and Mandarin technical terms.

Where the errors land

A 6% WER on a 60-minute call translates to approximately 360 incorrect or missing tokens. The relevant question is not the count, it is the distribution.

Finance vocabulary has properties that punish general ASR:

• Acronyms. TSMC, EBITDA, EBITDAX, ASIC, HBM, ARR, MNPI. Standard models often render these phonetically (DSMC, EBIT-DA spelled out, A6) or substitute homophones.
• Tickers and company names. AMD versus AMD Inc, ASML versus AMSL, Arista versus a rest, Vertiv versus Vertive. The error is small in edit distance, fatal in meaning.
• Numbers and units. 6% versus 60%, 200bps versus 2%, USD 5M versus 5 million (no currency). Currency specification is one of the failure modes that an analyst will not catch by skim-reading.
• Code-switched technical terms. Mandarin, Korean, German, and Japanese terminology that recurs in semis, autos, and pharma calls.

If 50 to 80 of the 360 errors land on these tokens, the transcript is not a 94% accurate document. It is a document where the analyst has to re-listen to every passage that contains a number, name, or acronym, which is most of the call.

What this costs the buyer

The direct cost is the call fee, USD 1,000 to USD 1,500. The hidden cost is the analyst time spent re-listening to the audio to verify material claims, which is the work the transcript was meant to eliminate. On a 60-minute call, an analyst working under IC deadline can lose 30 to 60 minutes to verification, on top of the 4 to 12 hour transcript delivery delay.

There is also a quieter cost. An analyst who finds two material errors in a transcript stops trusting the rest of it. The document moves from primary research artifact to rough notes, which means the next call gets transcribed again from the audio anyway. The transcription product gets bypassed even when the price has already been paid.

What to watch

The interesting signal will come from how AlphaSense, Tegus (now part of AlphaSense), and the in-house transcription teams at the major expert networks position their accuracy claims over the next 12 months. If they continue to quote conversational WER, the gap stays open. If they start publishing finance-specific accuracy on tickers, acronyms, and numerical tokens, the benchmark conversation shifts, and the buyer finally has a number that matches the use case.