What Is 911 Call Analysis, and Why Prosecutors Are Losing It

Your client's own 911 call may be the strongest piece of evidence against them, and the method prosecutors use to interpret it has a foundation that researchers, defense organizations, and even a state law enforcement agency have called into question.

The technique at the center of this is former Deputy Chief Tracy Harpster's "911 COPS Scale," which claims to detect guilt and deception from a caller's word choice, cadence, and grammar. According to a defense attorney's review of the underlying science, the theory rests on 20 variables and only two studies of roughly 100 calls each, about 200 cases total, and no one has been able to replicate Harpster's findings. Questions about whether calls were randomly selected, whether analysts knew case outcomes before listening, and why only English speakers were included remain unanswered.

The scale of the problem is not hypothetical. Investigative reporting has documented Harpster's methods in more than 100 cases across 26 states, with judges and prosecutors now pushing for bans and post-conviction review. The pushback comes from unusual corners:

• NACDL president Nellie King placed 911 call analysis on "the junk pile of fake scientific theories" and called the business behind it "dangerous and insidious."
• The Colorado Bureau of Investigation's chief of staff said the technique should never be "a deciding factor of an arrest or conviction."
• The University of Michigan's STPP policy brief found the practice lacks robust empirical support and has contributed to wrongful convictions.
• North Carolina's Office of Indigent Defense Services issued a statewide warning to defense attorneys to watch for it.

There is a separate, peer-reviewed body of linguistic research on 911 calls, but it points the opposite direction from Harpster's claims. A study in Security Informatics found that truthful callers actually display more negative emotion and anxiety and provide more detail, while deceivers use more negation and assent words. The same study notes human deception detection sits near 54%, only slightly better than chance.

That gap matters in practice because a prosecution narrative that your client "sounded deceptive" can be met with published research showing the opposite marker. If you're preparing a Daubert or Frye challenge, or just trying to get ahead of how the state will frame the call, start with your own grounded 911 call transcript analysis rather than a prosecution-supplied summary. The same discipline applies to recorded jail call transcription, where a client's own words often surface again in a different setting.

How to Challenge 911 Call Analysis Testimony: The Daubert/Frye File

When the prosecution's case leans on a "911 call analysis" expert, your job is not to out-analyze him. It is to show the court there is nothing scientific to analyze. The technique's documented weaknesses hand you a ready-made Daubert/Frye file.

Start with the research base. Harpster's theory rests on two studies of roughly 100 calls each, about 200 cases total, and according to defense-side analysis of the underlying science, no one has replicated his findings. That source also flags unanswered questions about random selection, analyst blinding, and English-only datasets. Harpster's paid 8-hour training course, sold to police departments nationwide, is a conflict angle worth raising on cross.

Then bring the institutional record. The University of Michigan STPP brief concludes the technique lacks robust empirical support and has contributed to wrongful convictions. Reporting documents its use in 100+ cases across 26 states. Build your challenge file around these resources:

• North Carolina IDS statewide warning flagging 911 call analysis for defense scrutiny
• NACDL's published treatment of the technique as forensic junk
• Exoneration Project and Center for Integrity in Forensic Sciences involvement in challenged cases
• Fair and Just Prosecution's call for post-conviction review

The peer-reviewed linguistics cuts the other way. A LIWC study of 911 homicide calls found truthful callers display more anxiety and detail, while deceivers use more negation and assent words. Human deception detection sits near 54%, barely better than a coin flip. When a prosecutor says your client "sounded guilty," that literature is your rebuttal.

Independently review the underlying audio. Dispatch-stage AI now generates transcripts and keyword flags automatically, so never rely on a prosecution summary. Run your own grounded, timestamped 911 call transcript analysis, then cross-check it against bodycam and reports. The same discipline applies to jail call transcription, where a client's recorded words often get introduced with the same false confidence. Every AI answer in Evidence Copilot cites the exact moment it came from and is marked for your review, so nothing enters a motion without human verification.

The New Wrinkle: AI-Generated 911 Transcripts in Discovery

By 2026, the 911 call analysis in your discovery packet might not have been written by a human at all. Dispatch-stage AI now generates machine transcripts, keyword flags, and call summaries before anyone dials a detective, and those artifacts can land in discovery without the traditional authentication trail you'd expect from a court reporter or a certified transcript.

Two vendors dominate this space. RapidSOS's Harmony AI provides real-time transcription supporting over 190 languages with automatic translation in over 130 languages and dialects, plus keyword detection for terms like "apartment," "fire," and "screams." BEI's January 2026 writeup describes AI transcription that "directly improves the seconds-to-dispatch metric" alongside speech analytics including keyword detection, vocal stress analysis, and contextual prioritization.

That's useful for dispatchers under load. It's a problem for you.

Machine transcripts mishear names, addresses, and negations. Keyword detectors miss context. Vocal stress analysis is a black box with no published error rate. And a call summary generated in the first eight seconds of a call is not a record of what the caller actually said.

Don't rely on prosecution-supplied summaries or AI keyword analyses. Treat them as leads, not evidence. Here's the workflow:

• Pull the original 911 audio file, not the PDF summary. If the State produced only a transcript, demand the audio under your discovery statute.
• Produce your own timestamped transcript with speaker diarization so you can cite the exact second a word was spoken.
• Compare your transcript line by line against what the State handed you. Mark every divergence.
• Check the AI keyword flags against the audio. Note every instance where the flagged word is absent, or where context flips its meaning.
• Log the chain: who generated the transcript, which model, which version, and when. If the State can't answer, that's a cross-examination point.

A plaintiff-side vendor marketing the same class of tools acknowledges that AI-generated transcripts "can be used to support legal arguments, but attorneys must ensure compliance with court rules and evidence authentication standards." Expect the other side to introduce these artifacts. Make them authenticate.

Grounded 911 call transcript analysis means every claim you make about the call ties to a timestamp in the original audio. That's how you impeach a machine transcript, and it's how you catch the moment the AI summary quietly dropped the caller's plea for help. If the same call is cross-referenced against bodycam or jail calls in the case, the same principle applies: jail call transcription should carry word-level timestamps too.

Evidence Copilot's 911 call transcript analysis workflow is built around this comparison: upload the original audio, get a diarized timestamped transcript, and diff it against whatever the State produced. Every AI output is marked for human review, and every citation points to the exact second in the audio. Book a 20-minute walkthrough at evidencecopilot.com/book if you want to see the diff view on a real call.

Your 911 Call Review Workflow, Step by Step

A 911 call lands in discovery as a single audio file, but the work it takes to turn that file into a usable defense tool is anything but simple. The right workflow catches contradictions before the prosecution does, flags language that the junk-science witnesses will quote, and gives you a transcript you can actually cite at the hearing.

Here's the step-by-step:

• Get the original audio. Request the native recording from the PSAP or law-enforcement evidence custodian, not a summary transcript. Vendor-produced AI transcripts from systems like RapidSOS or BEI's dispatch analytics may contain transcription errors that you cannot challenge without the underlying audio.
• Build a timestamped, diarized transcript. Speaker labels (caller, dispatcher, second caller, background voices) with word-level timestamps let you cite the exact second of any statement. When you rename a speaker once, every segment across the case updates.
• Cross-reference against the rest of discovery.
• Document each inconsistency with a citation.
• Flag the call against peer-reviewed research, not Harpster's indicators.

That third step is where most defenses get burned. The 911 caller's initial account, the police report's narrative, the bodycam audio, and any later witness statements should all be compared side by side. An early 911 account that differs from later testimony cuts both ways: it can impeach a prosecution witness, or it can corroborate your client's account of what actually happened.

When you find a contradiction, write it down with the exact timestamp in each source. "Caller says 'he ran out the back door' at 02:14 in the 911 audio; officer writes 'suspect fled on foot eastbound' in report page 3" is the kind of note that survives a cross-examination. In a workspace like Evidence Copilot, every new file gets compared against earlier evidence on ingest, and candidate report-vs-call inconsistencies surface as review signals before you have to ask. The same workflow extends to jail call transcription for cases with recorded custodial calls.

The fifth step matters because the prosecution's expert will likely lean on the 20-variable Harpster framework, which rests on only two studies of roughly 100 calls each and has not been independently replicated. Instead, anchor your objections in the 2014 peer-reviewed LIWC analysis, which found that truthful callers actually display more negative emotion and provide more detail, while deceivers use more negation and assent words. That finding directly undercuts the prosecution narrative that an anxious, detail-heavy caller is "suspicious."

For a deeper look at how independent 911 call transcript analysis fits into the broader evidence-review workflow, the rest of this guide walks through the Daubert challenge, the cross-exam questions, and the redaction steps for discovery turnover.