What 911 Call Analysis Is and Why It's Still in Your Cases

A detective tells the jury your client opened the 911 call with "hi," and that politeness itself is a sign of guilt. That's not a hypothetical. It's how 911 call analysis has entered trials across the country, and it's still showing up in 2026 cases.

The method traces to Tracy Harpster, a former Ohio police lieutenant whose 2006 master's thesis analyzed just 100 homicide calls and identified 16 "guilt indicators." The count grew to 18 in a 2009 paper and to 38 in his 2017 book, all built on an expanded sample of roughly 200 calls, according to a University of Cincinnati Law Review analysis. No one has replicated the findings. At least eight replication and validation attempts have failed, per the Innocence Project.

The scale of use is the problem. A ProPublica investigation documented more than 100 cases across 26 states where law enforcement applied Harpster's methods, and departments still pay up to $3,500 for about eight hours of training.

Courts have repeatedly rejected the technique as expert testimony. An Alabama court refused to qualify Harpster in 2009. A Nevada judge rejected a detective trained in the method. And a Michigan appellate court overturned the conviction of 16-year-old Riley Spitler, holding that 911 call analysis and related statement-analysis techniques should never have been admitted.

So prosecutors adapted. Instead of proffering an expert, they put trained detectives and dispatchers on the stand to walk juries through the audio and transcript as lay witnesses. The jurors hear guilt inferences drawn from word choice and tone without ever learning the disputed methodology behind them. Cornell Law Review scholarship calls this pattern "shadow forensics": the methods are never disclosed and the technique's founders are never offered for Daubert or Frye scrutiny.

The real-world stakes are documented:

• A Washington mother whose call began with "hi" was flagged as "too polite," her baby's death was reclassified as a homicide, and she entered an Alford plea to manslaughter.
• Huwe Burton, convicted at 16 after an officer testified he "did not seem emotional enough," served more than 20 years before his 2019 exoneration.
• Robert Roberson's flat affect was read as guilt. He was later diagnosed with autism spectrum disorder and has spent 20+ years on Texas death row, narrowly avoiding execution three times.

Dr. Jessica Salerno's research at Cornell explains why these readings are unreliable: both civilians and police become significantly more suspicious of low-emotion callers, and follow-up studies using real calls with known outcomes found the same association regardless of whether the caller was later convicted or exonerated. The suspicion tracks the listener's bias, not the caller's guilt.

For defenders, the takeaway is simple: assume the analysis is in your case until you've checked. When it arrives through lay testimony, your cross needs word-level impeachment, which means precise, timestamped transcripts of both the 911 audio and the officer's testimony about it. A grounded review workspace like Evidence Copilot supports that workflow, from 911 call transcript analysis to jail call transcription, with citations to the exact source moment so every quote you put to the witness is verifiable.

The Lay-Testimony Workaround: How 'Shadow Forensics' Slips Past Daubert

You won't see "Harpster" anywhere in the witness list. That's the point. Prosecutors have largely stopped proffering 911 call analysis as expert testimony because courts keep rejecting it under Daubert and Frye. Instead, they put a trained detective or dispatcher on the stand as a lay witness and have that officer walk the jury through the audio and transcript, flagging word choice, tone, and pauses as the call plays. Jurors hear "guilty" inferences without ever learning the disputed methodology behind them.

Cornell Law Review scholarship calls this pattern "shadow forensics": the underlying methods are never disclosed, the technique's founders are never proffered as experts, and the analysis largely escapes scientific evidentiary standards altogether. Harpster himself advises prosecutors trained in the method to guide juries through the audio and transcript without seeking expert qualification, according to a University of Cincinnati Law Review analysis.

So how do you spot it in your own case? The telltale signs are consistent:

• The witness testifies about the caller's word choice or emotional level as evidence of guilt, framed as personal perception rather than a method.
• The officer mentions "training" in call analysis or statement analysis, often an 8-hour course costing up to $3,500 in taxpayer funds, while denying any specialized expertise.
• Testimony tracks the known indicator list: commenting on the victim's eyes, asking whether to touch the body, being "too polite," or failing to plead for help.
• The witness uses the proprietary COPS Scale vocabulary (offender probability, passive defenses, resistance to the dispatcher) without attributing it to anyone.

The stakes are real. ProPublica documented more than 100 cases across 26 states where law enforcement used the method, which rests on a foundational study of just 100 calls and has failed at least eight replication attempts, according to the Innocence Project.

Practically, you need precise, citable transcripts of both the 911 audio and the officer's testimony about it. That's where grounded review work pays off: at Evidence Copilot, timestamped, diarized transcripts with word-level citations let you pin the officer's claimed "indicator" to the exact second of the call, and the same workflow covers related custodial statements through jail call transcription. When the witness says the caller "never asked for help," you want the transcript window that proves or breaks that claim, not a memory of the audio.

Once you can name the pattern, you can fight it.

Building the Challenge: Replication Failures, Court Rejections, and Cross-Exam Ammunition

Build the challenge before you need it, because 911 call analysis rarely announces itself as a scientific claim. It arrives through a detective or dispatcher who simply walks the jury through the audio, word by word, and lets the inference land without ever seeking expert qualification.

That workaround is the whole ballgame. Prosecutors trained in the method can guide jurors through both the recording and the transcript, which lets guilt inferences from tone or word choice reach the jury while the underlying methodology stays out of sight. Cornell Law Review scholarship calls this pattern "shadow forensics": the methods are never disclosed and the technique's founder is never proffered as an expert, so the analysis largely escapes Daubert or Frye review. Your challenge file has to be assembled before that testimony shows up, not after.

Here's what belongs in it.

• The foundational study: Harpster's 2006 master's thesis, built on a sample of just 100 calls, later expanded to 200 for the 2017 book, with the indicator count growing from 16 to 18 to 38 along the way.
• The replication record: at least eight failed replication or validation attempts, and no one has been able to reproduce Harpster's findings.
• Prior court rejections: an Alabama court rejected Harpster as an expert in 2009, a Nevada judge rejected a detective trained in the method, and a Michigan conviction was overturned because the technique should not have been admitted.
• The disclosure gap: Harpster has never testified as an expert in a criminal trial and has never made his sample data public or explained how he obtained it.

The indicator list itself is your cross-exam ammunition. Saying "bleeding" reads as innocent while saying "blood" reads as guilty; opening a call with "hi" can be scored as too polite, and therefore suspicious. There are no published guidelines accounting for anxiety, education level, cultural factors, speech impediments, or mental health, which makes the scoring system look arbitrary the moment you walk a witness through it.

Then bring in Salerno's emotion research. Both civilians and police officers become more suspicious when callers express low emotion, and follow-up work on real 911 calls found that low-emotion-equals-guilt association held regardless of whether the caller was later convicted or exonerated. That's empirical impeachment of the core assumption, not just a methodological quibble.

Anticipate the rebuttal. Harpster claims he has examined more than 2,500 equivocal death cases and is unaware of any innocent person convicted through the analysis, and he credits the method with helping solve over 1,500 homicides. Set that against the Michigan overturn and the documented wrongful convictions, including Huwe Burton's 20-plus years before exoneration in 2019.

Word-by-word impeachment needs word-by-word records. Evidence Copilot's 911 call transcript analysis produces timestamped, diarized transcripts that cite the exact moment a phrase was spoken, so when a witness attributes guilt to a single word, you can pull the audio and the transcript line to the second.

The Grounded Transcript Workflow: Turning 911 Audio into a Citable Record

The problem with 911 call analysis isn't just that the science is thin. It's that prosecutors rarely present it as science at all. Per legal scholarship, Harpster advises trained prosecutors to "guide juries through both the audio and transcript of the calls" without ever seeking expert qualification, so the guilt inferences arrive through lay testimony that never faces a Daubert or Frye hearing.

That means your defense work starts at the discovery screen. Flag any testimony that ties a caller's word choice, tone, or emotional level to a guilt inference, whether it comes from a detective, a dispatcher, or an officer who took an eight-hour course. If the witness can't explain the methodology behind the inference, that's the shadow-forensics pattern Cornell's Christopher Lau describes: the methods stay undisclosed and the technique's founder is never proffered as an expert, which lets the analysis escape evidentiary challenge.

Once you've flagged a call, build a word-by-word impeachment record. You need a timestamped, diarized transcript of the 911 audio so that when the officer testifies the caller "never asked for help" or "sounded too calm," you can quote the exact second of audio that contradicts him. Evidence Copilot's 911 call transcript analysis workflow produces those grounded transcripts, with every AI answer citing the exact source moment and marked for human review before anything goes in a brief.

The same grounded approach covers the rest of the custodial-statement record. Jail call transcription with word-level timestamps lets you line up what your client said on the phone against what the report claims was said, and auto-signals surface statement-versus-call-audio inconsistencies before you'd catch them manually.

Your cross-examination prep should target the indicators themselves, and the transcript gives you the ammunition:

• "Bleeding" is supposedly innocent while "blood" is supposedly guilty, a distinction with no empirical basis.
• A Washington mother who opened her call with "hi" was flagged as "too polite," and her baby's death was reclassified as homicide.
• The indicator list grew from 16 (2006 thesis) to 18 (2009) to 38 (2017 book), all based on roughly 200 calls.
• No guidelines account for anxiety, education, culture, speech impediments, or mental health.

Dr. Jessica Salerno's research at Cornell found that both civilians and police grow more suspicious when callers express low emotion, and follow-up studies using real 911 calls found the same association regardless of whether the caller was later convicted or exonerated. That's empirical impeachment you can put in front of a court, backed by the broader record: a foundational study of only 100 calls and at least eight failed replication attempts, documented by the Innocence Project.

Keep the originals immutable and every excerpt cited to its source moment. When the state's theory rests on what a caller said and how they said it, your record should be able to prove exactly what was said, second by second.