What Is 911 Call Analysis, and Why It Shows Up in Your Discovery

If a dispatcher or detective starts talking about your client's 911 call "sounding staged," you're not listening to an observation. You're listening to the output of a technique most defenders have never heard of, and it may already be in your discovery.

911 call analysis claims to detect guilt from a caller's word choice, pauses, and tone. It started with Tracy Harpster, an Ohio deputy police chief with no scientific background, whose 2006 master's thesis analyzed just 100 homicide calls and identified 16 "guilt indicators." The list kept growing: 18 indicators in a 2009 paper with former FBI co-authors, then 38 in a 2017 book based on 200 calls. The companion tool, the COPS Scale, sorts callers across six categories, from who the call is about to "additional guilty indicators."

Here's the problem: the method has never validated. ProPublica's investigation found zero of five validation studies (including two by the FBI's Behavioral Analysis Unit) supported the technique, yet documented more than 100 cases across 26 states where it influenced arrests, prosecutions, and convictions. Harpster charges up to $3,500 per eight-hour training class, typically paid with tax dollars, and police departments keep signing up.

The scale is worse than the case count suggests. The FBI circulated an unlabeled version of the 2008 study to roughly 200,000 bulletin readers before peer review, accelerating adoption nationwide. Harpster claims consultations on 1,500-plus homicide investigations, and prosecutors have learned to present the analysis without naming it.

That's the part that should worry you: it slips past case screening. A Virginia public defender put it bluntly: "I have never heard of any of that claptrap in my jurisdiction." Because the underlying methods are never disclosed and Harpster has never testified as an expert, the technique "largely escapes Daubert or Frye evidentiary challenge," arriving instead as lay opinion from an officer who took a class. Cornell's law review calls it shadow forensics for exactly that reason.

So when you review a homicide file, treat these as red flags:

• Any dispatcher or officer commentary on the caller's "tone," "staging," or word choice
• Testimony about "guilty indicators" like saying "somebody" instead of a name, or being "inappropriately polite"
• A training certificate referencing Harpster, the COPS Scale, or 911 call analysis
• Prosecution exhibits walking the jury through call audio and transcripts

Screening means actually listening to the call with accurate, timestamped transcripts in hand, alongside the reports and bodycam. That's the workflow we built Evidence Copilot around: grounded 911 call transcript analysis and jail call transcription, with every AI answer cited to the exact second of audio and marked for human review. You can't challenge what you haven't identified.

The Shadow Forensics Problem: How It Evades Daubert and Frye

The trap in 911 call analysis cases isn't weak science alone. It's that the science never shows up in court under that label. Tracy Harpster has never testified as an expert in a criminal trial despite consulting on over 2,500 homicide investigations, and he instead advises prosecutors to walk juries through the audio and transcripts without ever seeking expert qualification, according to the University of Cincinnati Law Review.

The Cornell Law Review's shadow-forensics framework names the mechanism precisely: because the underlying methods are never disclosed and the founders are never proffered as experts, the technique "largely escapes Daubert or Frye evidentiary challenge." Instead, a dispatcher or officer who attended a Harpster training takes the stand and offers a guilt opinion as ordinary perception, framed as lay testimony under FRE 701 rather than expert testimony under FRE 702. The framework's core argument is that a demeanor-based guilt opinion derived from a taught methodology is expert testimony in disguise, and the disguise is the point.

Prosecutors have said so in writing. Missouri prosecutor Leah Askey wrote in a 2016 email that she sidestepped required admissibility hearings by "getting creative … without calling it 'science,'" adding, "Of course this line of research is not 'recognized' as a science in our state," per the ProPublica investigation republished by Forensic Magazine.

A standard Daubert motion won't catch this on its own because there's no proffered expert to challenge. You need a paired attack:

• A motion in limine arguing the guilt opinion is surreptitious expert testimony under FRE 701/702, using the shadow-forensics framework.
• A Daubert/Frye record built from the five failed validation studies, including two FBI BAU replications, ready for the moment the state claims the method has any scientific footing.
• Voir dire of the testifying dispatcher or officer on their Harpster training, since the training itself is the undisclosed methodology.

The Russ Faria case shows the pattern in practice: a dispatch supervisor trained in Harpster's class testified his call "sounded staged," and Faria spent 3.5 years in prison before an acquittal on retrial. ProPublica has documented more than 100 cases across 26 states, described as likely a fraction of the real figure.

Practically, your first move is finding the contamination early. A 911 call transcript analysis with word-level timestamps lets you compare the call against the police report and bodycam before the preliminary hearing, and jail call transcription catches the same word-choice suspicion when it migrates to recorded calls. If a report describes your client's "tone" or "staging," you've found the shadow before the state calls its witness.

The Science Against It: Zero-for-Five Validation and the FBI's Own Caution

When you put the empirical record on the table, 911 call analysis doesn't just look weak. It looks like it has never once passed a validation study.

Start with the scorecard. Five studies have attempted to validate the method, including two from the FBI's Behavioral Analysis Unit plus research out of Villanova and James Madison University, and none succeeded, according to a ProPublica investigation. At least six peer-reviewed studies have also failed to replicate the training's findings, per Cornell's coverage of the Salerno research. That's the peer-review and error-rate failure you need for a Daubert record.

The reliability numbers are just as useful for cross. In the FBI's own data, seasoned investigators with 10+ years of experience agreed on whether a caller showed voice modulation in only 72% of 175 calls. And the flagship "guilty" indicator collapses under scrutiny: over 80% of both innocent and guilty callers used the word "just," with involved callers averaging 4 uses per call against 2.5 for uninvolved. That's an overlap problem, not a discriminator.

The FBI's own authors say so. The FBI Law Enforcement Bulletin concludes that voice modulation and minimization indicators "should not be used on their own for evaluating veracity versus deception," and warns that investigators assessing minimization "may run the risk of having a 'confirmation bias.'" FBI co-author John Jarvis went further, saying he doesn't see where the original work "rises to the level of success by the scientific community."

One nuance to anticipate: prosecutors sometimes cite FBI research selectively, pointing to the Bulletin's framing of some indicators (like "just" frequency or location statements) as weak-to-moderate signals worth further study. Your answer is the Bulletin's own caveat, in its own words, plus its recommended best practice: compare call statements against case facts and interviews rather than analyzing calls in isolation.

Expect the creators' counterpoints too. Harpster denies any innocent person was convicted based on the analysis and claims it has exonerated people. Susan Adams argues "no single indicator can be used to determine the likelihood of innocence or guilt" because the method examines indicators in combination. Both concessions cut your way: Adams's own statement concedes no single indicator works, which is exactly how the technique gets applied on the witness stand.

For motion practice and cross-exam prep, this is where a grounded review workspace earns its keep. When you're pulling word-level timestamps from the 911 audio and mapping every claimed "indicator" against the 911 call transcript analysis record and the case file, having each citation tied to the exact moment in the call is what turns these statistics into exam questions. The same discipline applies when you're working jail call transcription into the comparative record the FBI itself recommends.

The Defense Playbook: Motions, Cross-Exam, and Case Law to Cite

Start by treating every 911 call in discovery as a potential contamination site. Given that 95% of surveyed homicide investigators often or always review 911 calls, and the technique has surfaced in more than 100 cases across 26 states according to ProPublica's investigation, any dispatcher or officer commentary about a caller's "tone," "staging," or word choice deserves a hard look. Many defenders have never heard of the method, which is exactly why it slips through.

Build three concrete work products:

• A discovery screen: flag every report, supplement, or transcript where an officer characterizes demeanor ("sounded scripted," "inappropriately polite," "no urgency").
• A motion in limine and FRE 701/702 challenge: frame demeanor-based guilt opinions as disguised expert testimony, the core argument in Cornell Law Review's "shadow forensics" framework.
• A Daubert record: attach the failed validation studies, the FBI's own cautions, and the inter-rater reliability data.

The empirical record does the heavy lifting. Zero of five independent studies validated the method, and at least six peer-reviewed studies failed to replicate the training's findings. Even seasoned investigators with 10+ years on the job agreed on whether callers showed voice modulation in only 72% of 175 calls, per the FBI Law Enforcement Bulletin. That figure alone undermines any officer who claims to reliably detect guilt from a caller's tone.

Cross-examination writes itself with the case examples. In Washington, a detective flagged a mother's first word, "hi," as suspicious, and she took an Alford plea to manslaughter. A Michigan caller landed on a suspect list partly for saying "I need help." And in Fremont, Ohio, a detective elicited a confession after throwing down a 911 call analysis handout. Russ Faria spent 3.5 years in prison before acquittal on retrial, after a dispatch supervisor trained in the method testified his call "sounded staged," and Riley Spitler's Michigan conviction was overturned when the technique was improperly admitted.

Anticipate the counterpunch. Harpster denies any innocent person was convicted through the analysis, and co-author Susan Adams insists no single indicator determines guilt. Pin them to the reliability data and the failed replications.

Grounded review starts with accurate transcripts. Pairing timestamped, diarized 911 audio and jail call transcription against reports and bodycam lets you compare statements to case facts rather than analyzing calls in isolation, the best practice the FBI itself recommends. Evidence Copilot's 911 call transcript analysis keeps every citation tied to the exact moment, with all AI output marked for your review.

Grounding the Analysis: Transcript Work That Survives Scrutiny

The FBI's own Behavioral Analysis Unit researchers put it plainly: 911 call statements should be compared against case facts, written statements, and interviews rather than analyzed in isolation, and even the linguistic indicators they studied should not be used on their own for judging veracity. That's the standard you should hold your own 911 call analysis to, and it's also the standard that exposes the prosecution's shortcut. Harpster's method, by contrast, was built from a single 100-call master's thesis and has failed validation in every independent test, including two FBI studies.

Doing the comparison the FBI recommends takes more than a good ear. You need accurate, timestamped, speaker-labeled transcripts of the 911 audio sitting next to the bodycam footage, the police report, and any written statements. If a detective claims your client's call "sounded staged," you can't cross-examine that claim without knowing exactly what was said, when, and by whom. The same discipline applies to jail calls, where a client's casual remark can be matched against what he told police three days earlier.

That's where a grounded workspace earns its keep. Evidence Copilot's 911 call transcript analysis produces diarized transcripts with word-level timestamps, so every disputed phrase pins to the exact second of audio. The jail call transcription workflow does the same for recorded calls, and every AI answer carries a clickable citation into the source moment rather than a paraphrase you have to trust.

• Auto-signals on ingest: each new file is compared against evidence already in the case, so report-vs-call inconsistencies surface as review signals before anyone thinks to ask.
• Citation-backed answers: when you ask what the caller said about the weapon, the answer quotes the transcript and links the timestamp, not a summary.
• Human review by design: every AI output is marked as generated and pending attorney review, and originals stay immutable.

The reliability numbers argue for this grounded approach. Even seasoned investigators with 10+ years of experience agreed on whether a caller demonstrated voice modulation in only 72% of 175 calls in the FBI's own research. And Cornell's Salerno team found that identical scripts attracted different suspicion levels depending on emotional tone, with officers reading a calm report of a violent event as more suspicious. When trained raters can't agree, a transcript tied to verifiable case facts is the only thing that survives cross.

If you want to see the workflow on your own discovery, Evidence Copilot publishes its pricing and offers a 20-minute walkthrough at evidencecopilot.com/book.