Speech recognition models have closed most of the accuracy gap on conversational audio in the last 24 months. OpenAI's Whisper, Deepgram's Nova-3, AssemblyAI's Universal-2, and Google's Chirp family all report sub-5% word error rates on standard benchmarks. None of them have displaced human reviewers from the financial transcription workflows that touch the investment process. The reason is not that the models are bad. It is that WER on a conversational benchmark is the wrong metric for a transcript that has to survive an investment committee.

INFLXD's view: pure-AI transcription is a price floor, not a product. The structurally durable position is AI-first with a human final pass, and the economics of that position only get better as the models improve.

What the WER number is hiding

A 5% word error rate sounds like solved-problem territory until you look at which 5% breaks. Conversational benchmarks are dominated by everyday vocabulary. Financial audio is dominated by tickers, segment names, non-GAAP metrics, foreign-listed entities, and acronyms that the training distribution has barely seen. "TSMC" gets mistranscribed as "DSMC." "EBITDAX" collapses to "EBITDA." A guidance number stated as "three to five percent" gets rendered as "3.5 percent," which is a different number. None of these errors register on standard WER scoring against a generic reference transcript. All of them matter when an analyst is building a model.

The second failure mode is diarization. Earnings calls routinely run six to ten speakers (operator, CEO, CFO, plus four to seven sell-side analysts in Q&A). Expert calls and industry panels add more. Deepgram's own documentation acknowledges that diarization accuracy degrades with speaker count and acoustic similarity. The buy-side specifically pays for transcripts of multi-party calls, so the workloads where diarization is hardest are the workloads that matter most.

MNPI redaction is not an ASR problem

Material non-public information does not announce itself in the audio. A speaker mentioning an internal forecast that contradicts public guidance is producing text that an ASR system will transcribe perfectly and that a compliance team has to flag and redact before delivery. The judgment of what is material and what is non-public is human. Recent SEC enforcement actions on inaccurate transcript-derived filings have made clear that "the model said so" is not a defense.

Expert network compliance teams strip MNPI from call transcripts before they reach clients. That work is unautomatable today, not because the language models cannot read, but because the legal exposure of getting it wrong sits with a human reviewer and a compliance officer, not with a vendor's API.

The economics actually favor the hybrid

Pure-AI transcription runs roughly USD 0.01 to 0.02 per audio minute at scale. Human-reviewed pipelines run USD 0.20 to 0.40 per minute depending on turnaround and accuracy guarantees. The naive read is that AI is 20x cheaper. The real read is that the cost of a single bad transcript inside an IC memo, a 13F, or an SEC filing dwarfs the entire annual transcription budget of most funds.

Aiera and Quartr have both moved toward human-final-pass pipelines for their financial customers, framing it as a quality guarantee rather than a cost line. That is the right framing. The customer is not buying minutes of audio converted to text. They are buying defensibility.

Where this goes next

We see three paths:

Bull case for pure-AI. Foundation models trained on financial audio specifically (earnings calls, expert calls, conference panels) close the vocabulary and diarization gaps. Compliance gets partially automated by LLM-based redaction with human spot-checks. The human pass shrinks from full review to exception handling. Cost per minute drops toward USD 0.05.

Base case (most likely). AI handles the first pass at near-zero marginal cost. Humans review every transcript that touches the investment process, with the human's job shifting from typing to judgment: catching ticker errors, fixing diarization, flagging MNPI, attesting to accuracy. Cost stays in the USD 0.15 to 0.30 range. This is the model Aiera and Quartr are already operating.

Bear case for pure-AI. A high-profile SEC enforcement action against a fund that relied on AI-only transcripts hardens the regulatory posture. Human review becomes a compliance requirement, not a quality choice. Pricing power shifts to vendors with auditable human-in-the-loop workflows.

What to watch

The metric that matters is not WER. It is the rate at which transcripts produced by a given pipeline survive a downstream audit (IC memo, 13F citation, SEC filing) without correction. No vendor publishes that number yet. The first one that does, with credible methodology, will reset how the buy-side evaluates this category.