Counseling research depends on behavioral coding: trained people listen to a therapy session and label each counselor utterance as a question, a simple reflection, a complex reflection, a piece of advice, and so on. Supervision, fidelity checks and outcome studies all rest on these labels, and producing them costs 7 to 10 hours of human work per hour of session. The obvious fix is to run speech recognition on the recording and let a classifier assign the codes. That fix fails badly. On real Motivational Interviewing sessions, a coder trained on human transcripts falls from 0.694 to 0.528 accuracy when it reads automatic transcripts instead (Evaluating Automatic Speech Recognition Quality and Its Impact on Counselor Utterance Coding). The same study finds that clients are recognized worse than therapists, and that emotionally loaded content words are often lost ("I really felt like I was dying" became "I really felt like I was there").
We argue that this drop is not uniform noise. It lands hardest on reflections, the code at the center of Motivational Interviewing fidelity, and there is a structural reason. A reflection is recognized by what it shares with the client's previous turn: the counselor restates the client's content, sometimes adding meaning. When the recognizer mangles the client's words, that shared content disappears from the text, even if the counselor's own turn was transcribed perfectly. Our proposal is to detect the echo in a space of concepts rather than surface words. Content mentions in both speakers' turns (substances, symptoms, medications, and the everyday life topics clients raise) are mapped to shared concept identifiers by a normalizer. The normalizer is trained to tolerate recognition errors by combining psychotherapy-specific word embeddings with a representation of how words sound. Coding then depends on whether the counselor echoed the client's concept, not on whether the two transcripts happen to share a string.
Core Claim
The claim has two parts, and it can be falsified. The first part is diagnostic: most of the clean-to-automatic accuracy gap in counselor coding comes from reflections, and specifically from exchanges where the client's turn was misrecognized. The second part is constructive: a feature that measures concept-level overlap between a counselor turn and the preceding client turns recovers a substantial share of that reflection-specific gap. That feature relies on entity normalization that tolerates spelling and sound confusions, and it weights each match by the recognizer's confidence. We expect it to beat simply feeding more (noisy) context to a strong classifier or a large language model.
The prior evidence motivating this is scattered but consistent. Early reflection detectors modeled the dependence of a reflection on the preceding client statement with a sequence model over turns (A Case Study: Detecting Counselor Reflections in Psychotherapy for Addictions using Linguistic Features), which means the relational signal was always part of the task. On automatic transcripts, adding two neighboring turns of context to BERT helps only slightly, lifting macro F1 from 0.2085 to 0.2290 (Evaluating Automatic Speech Recognition Quality and Its Impact on Counselor Utterance Coding). We read this as a sign that raw noisy context carries the relational signal poorly. Work on speech-based entity recognition points the other way from "more context": it found that recognizer confidence should be used to reject doubtful words rather than trust them, raising precision from 70.67% to 78.13% (Discriminative named entity recognition of speech data using speech recognition confidence). Our design reconciles the two. It uses context, but only through confidence-weighted concept matches.
What is new, relative to the closest work, is this. The ASR-and-coding study treats the classifier as a black box and stops at counting dictionary keywords. It does no normalization and uses no phonetic or domain representations. The phonetic entity-resolution work (Phonetic Embedding for ASR Robustness in Entity Resolution) handles only short voice-search queries against a movie or book catalog. Clinical normalization systems (CNN-based ranking for biomedical entity normalization) assume clean written notes. None of them uses normalization as a relational signal between two speakers in a conversation, and none targets a behavioral code.
Research Program
The work has three components, built in order so that each can fail on its own.
Spoken counseling mention corpus. No resource in our file annotates clinical or content entities in spoken therapy, so we build one. We take the Motivational Interviewing corpus of Building a Motivational Interviewing Dataset, which has 277 sessions and 97.8 hours of audio with codes under the Motivational Interviewing Treatment Integrity scheme (MITI 4.0). If audio access for that corpus is granted, we run a recognizer over it; otherwise we work with the 213-session spoken corpus of the ASR-quality study under a data agreement. We then force-align the human transcripts to the automatic ones. On a subset of roughly 40 sessions, annotators mark content mentions in client and counselor turns. These are linked to a concept inventory with two tiers. The first tier is a clinical subset of UMLS/SNOMED covering substances, symptoms and medications. The second is a smaller, curated inventory of recurring life-domain and affect concepts, such as family conflict, work, or feeling stuck; these are not in medical ontologies but are what reflections often echo. Annotators also mark which counselor reflections echo which client mention.
This annotation measures one of the proposal's main risks directly: what fraction of reflections have any concept anchor at all. If fewer than about half do, the concept pathway can only ever help a minority of reflections, and we would say so.
Error-tolerant normalizer. The normalizer follows the two-stage design of candidate generation followed by ranking. We take seriously the finding in CNN-based ranking for biomedical entity normalization that a two-stage system can never recover a concept that the first stage fails to propose. Candidate generation therefore looks up concepts by string match, by domain-embedding similarity, and by sound-alike similarity, so that a garbled mention can still retrieve its true concept.
The ranker combines two representations. The first is word embeddings pre-trained on psychotherapy transcripts, which Behavioral Coding of Therapist Language in Addiction Counseling Using Recurrent Neural Networks found to be the single largest source of gain for counselor coding. The second is a phonetic encoder in the style of Phonetic Embedding for ASR Robustness in Entity Resolution: a twin-network model that maps a word's pronunciation to a vector, so that words that sound alike land close together. We train it on confusion pairs harvested from the aligned human and automatic transcripts of training sessions. We take the recipe of mixing a small share of recognizer-derived pairs with clean pairs from that paper, which found training on only recognizer output hurt, and we tune the mixing ratio rather than fixing it. The normalizer can also output "no concept" for a mention, with a calibrated threshold, because many garbled tokens should map to nothing.
Concept-echo coder. For each counselor utterance, we compute a soft overlap between its linked concepts and those of the previous one to three client turns. Each match is weighted by the normalizer's confidence and by the recognizer's token confidence. This small set of echo features is given to a BERT coder alongside the utterance text. The design is deliberately narrow: the coder is standard, and only the relational evidence is new. We report results at the utterance level and also on session-level fidelity summaries such as the reflection-to-question ratio, because those summaries are what clinics actually use. Session-level human agreement is high (ICC 0.75 to 1.0 in Building a Motivational Interviewing Dataset), so they are a meaningful target.
Baselines
The naive reference point is a BERT coder trained and tested on automatic transcripts with no context. This mirrors the setup of Evaluating Automatic Speech Recognition Quality and Its Impact on Counselor Utterance Coding, which reports 0.694 accuracy on human transcripts falling to 0.528 on automatic ones, and a macro F1 of 0.2085 at zero context. We also report the same coder on human transcripts as the ceiling. All our gains are expressed as the fraction of the ceiling-to-naive gap that is recovered, overall and for simple and complex reflections separately.
The strongest discriminative competitor is that same coder with two turns of context, trained on a mix of human and automatic transcripts. The study above found that a 40% human-data mix recovers 80 to 85% of the gain from full human training, and that context lifts macro F1 to 0.2290. We run it on our splits with the same recognizer output, the same encoder, and the same hyperparameter search budget as our model. We sweep context lengths of 0 to 3 turns for both systems.
The strongest generative competitor is a large language model given the full coding manual, the conversation history, and the fine-to-coarse label mapping of Eliciting Motivational Interviewing Skill Codes in Psychotherapy with LLMs: A Bilingual Dataset and Analytical Study. On clean transcripts it reached macro F1 0.73 on BiMISC with the manual and 0.58 without. We found no reported result on Lacuna for this approach on automatic transcripts, so we measure it ourselves on identical sessions, both with and without our echo features written into the prompt.
Two ablations isolate the key ingredient. The first, which we call surface echo, replaces concept matching with plain word overlap and sentence-embedding similarity between the counselor turn and the client turns. This tests whether normalization adds anything beyond relational features in general. The second, called no-sound echo, removes the phonetic encoder and the recognizer-confidence weighting. This tests whether error tolerance, rather than normalization alone, carries the gain.
Beating the field means the following. Our model must significantly exceed the context-plus-mixing coder and the manual-prompted model on reflection F1 and on the ICC for the reflection-to-question ratio, using paired session-level bootstrap. It must also beat surface echo, and it must do so mainly in exchanges where client-turn error is high.
For the normalizer itself, the relevant comparison points are the 10-rule sieve and the CNN ranker of CNN-based ranking for biomedical entity normalization. On clean written text they report 84.65% and 86.10% accuracy on NCBI Disease, and the ranker reports 90.30% on ShARe/CLEF. We found no reported result on Lacuna for spoken counseling mentions, so we retrain both on our corpus under the same budget. We report accuracy and candidate recall at k on human-transcript mentions and on their recognizer-garbled counterparts.
Why not just…?
Why not just use a stronger recognizer and a manual-prompted large language model? This is a real threat. A modern recognizer may transcribe clients well enough, and a large language model with the full manual may infer echoes from noisy context, so that the gap largely closes. The evidence against this is indirect. Pretrained language models degrade steadily as word error rises (ASR-GLUE: A New Multi-task Benchmark for ASR-Robust Natural Language Understanding), and the 0.73 manual-prompted result was measured only on clean text. Rather than argue, we measure it. If the prompted model on automatic transcripts comes within noise of its own clean-transcript score on reflections, the premise that a gap exists fails and we would redirect toward the session-level and fairness analyses below. If a gap remains, we test whether putting the echo features into the prompt closes it, which would show the contribution holds even for large models.
Why not just train on automatic transcripts with more context? Data mixing already recovers most of the gain on aggregate accuracy, but aggregate accuracy is dominated by frequent, easy codes such as questions. The tiny change in macro F1 from adding context (0.2085 to 0.2290) suggests that noisy context carries relational signal poorly. The per-code comparison against our strongest discriminative baseline settles this. If reflection F1 under mixing plus context matches ours, the concept pathway is unnecessary.
Why not just correct the transcript first, or match with sentence embeddings? A general-purpose correction step or plain semantic similarity might restore the echo without any ontology. The surface-echo ablation tests the similarity route directly. A correction pass of the manual-prompted model, run before the coder, is added as a further baseline. Correction models trained on written text have no mechanism for recovering a content word replaced by a phonetically similar function word ("dying" to "there"), which is exactly where the phonetic candidate generator should help. That is where we expect the difference, and it is what the stratified analysis checks.
Proposed Experiments
Attribution experiment. This is the decisive test of the mechanism. Starting from human transcripts, we substitute the aligned automatic segments for only client turns, then for only counselor turns, and code each hybrid. The prediction is that reflection F1 falls more under client-only corruption, even though counselor text is untouched, and that the echo features recover most of the loss in that condition. If reflection F1 barely moves under client-only corruption, the relational account is wrong, and the concept pathway has no reason to help.
Main comparison. This covers the full automatic setting against all baselines. The primary metrics are per-code F1 for simple and complex reflections, and session-level ICC with human coders for the reflection-to-question ratio and the percentage of complex reflections. Results are stratified by client-turn error rate and by speaker gender and role, because the source study documents recognition disparities along those lines. A method that narrows the gap for high-error speakers has fairness value beyond its average accuracy.
Normalization check. On the annotated mentions, we compare accuracy and candidate recall at k for the sieve, the CNN ranker, our ranker without the phonetic encoder, and the full model. We report human-transcript and automatic-transcript conditions separately, plus calibration of the "no concept" decision.
Coverage check. We report the fraction of human-coded reflections that have a concept anchor and the recovery achieved within that subset. This makes explicit how much of reflection coding the approach can reach at best. If the anchored subset is small, that is itself a useful negative result: it shows that reflections in counseling largely echo content no fixed inventory captures, and it would point future work toward learned concept clusters instead of ontologies.
References
- Evaluating Automatic Speech Recognition Quality and Its Impact on Counselor Utterance Coding
- A Case Study: Detecting Counselor Reflections in Psychotherapy for Addictions using Linguistic Features
- Discriminative named entity recognition of speech data using speech recognition confidence
- Phonetic Embedding for ASR Robustness in Entity Resolution
- CNN-based ranking for biomedical entity normalization
- Building a Motivational Interviewing Dataset
- Behavioral Coding of Therapist Language in Addiction Counseling Using Recurrent Neural Networks
- Eliciting Motivational Interviewing Skill Codes in Psychotherapy with LLMs: A Bilingual Dataset and Analytical Study
- ASR-GLUE: A New Multi-task Benchmark for ASR-Robust Natural Language Understanding