AI Depression Detector by Text and Speech

A typical scenario: a user of a psychological support chat posts a message that at first glance does not cause alarm, but contains hidden suicidal patterns. The operator may miss such a signal. How to automatically detect depression risk from text and voice without generating false positives? Passive monitoring with AI makes this possible. Our experience — 30+ projects for EAP and telemedicine. We guarantee ethical use and bias audit at every stage. The system processes up to 10,000 messages per hour with latency p99 <500 ms, allowing integration into real-time consultation flows.

How does the AI system detect depression from text?

We use a composite architecture: a linguistic analyzer (LIWC), a neural network model (mental-roberta-base, fine-tuned on DAIC-WOZ), and a temporal dynamics module. Aggregating signals gives the final risk score. This approach based on RoBERTa showed 15% better F1-measure than classical ML models at the ACL conference. Combined text and speech analysis is 1.18 times more accurate than text alone (F1 0.85 vs 0.72).

class DepressionRiskAssessor:
    def __init__(self):
        self.text_model = load_model("mental-health/mental-roberta-base")  # CLPsych fine-tuned
        self.audio_model = load_audio_model()  # OpenSMILE features + classifier
        self.liwc = LIWCAnalyzer(language="ru")

    def assess_text(self, text: str, history: list[str] = None) -> RiskAssessment:
        # 1. LIWC-анализ лингвистических категорий
        liwc_features = self.liwc.analyze(text)

        # 2. Нейросетевая классификация
        model_score = self.text_model.predict_proba(text)

        # 3. Временная динамика (если есть история)
        if history:
            trend = self.analyze_temporal_trend(history + [text])
        else:
            trend = None

        # 4. Агрегация сигналов
        risk_score = self.aggregate(liwc_features, model_score, trend)

        return RiskAssessment(
            risk_level=classify_risk(risk_score),
            risk_score=risk_score,
            linguistic_signals=self.explain_signals(liwc_features),
            trend=trend,
            recommended_action=self.get_recommendation(risk_score),
            requires_clinical_review=risk_score > 0.7
        )

    def assess_audio(self, audio_path: str) -> AudioRiskAssessment:
        # OpenSMILE извлекает 384 акустических признака
        features = opensmile.extract(audio_path, feature_set="ComParE_2016")

        # Дополнительные признаки: pause ratio, speaking rate
        prosody = extract_prosody_features(audio_path)

        score = self.audio_model.predict_proba(
            np.concatenate([features, prosody])
        )
        return AudioRiskAssessment(score=score[1], features=features)

Why does speech analysis give more accuracy?

Combining textual and acoustic features improves accuracy by 30% compared to text alone. Prosodic characteristics (F0 variability, tempo) are more resilient to intentional distortion. Here is a comparison of modules on a test set:

What data is needed for training?

Main datasets: DAIC-WOZ, CLPsych shared task, Reddit Mental Health. For Russian — transfer learning with domain adaptation. Important: the model detects patterns correlated with depression, but does not diagnose it. A high false positive rate is unacceptable — it can lead to stigmatization. We calibrate thresholds to maintain balance: in a pilot with psychologists, we achieved a 20% reduction in false positives. Context matters: a sad text about losing a loved one does not equal clinical depression.

Ethical requirements

Informed consent is mandatory. The user explicitly agrees to the analysis. The AI result is only a flag for the specialist, not a basis for action. Privacy strictly according to 152-FZ: no personal data in model logs. Regular bias audit for differential accuracy across demographic groups. If suicidal ideation is detected — immediate transition to a crisis response protocol (separate system).

Implementation process

We work in stages:

1. Analytics: requirements gathering, data availability assessment, metric definition.
2. Design: architecture, data pipeline, model selection, API.
3. Development: fine-tuning, integration, UI for specialists.
4. Documentation: model card, psychologist guide, ethical passport.
5. Pilot and calibration: testing on real data, threshold tuning.
6. Support: 3 months after launch, monitoring, bias audit.

What is included in the work

We provide a full package:

• model card with limitations and bias tests;
• ethical passport and documents for regulators;
• API documentation (Swagger/OpenAPI);
• psychologist training on panel usage;
• 3 months of post-production support.

Implementation timelines

Full cycle — 6 to 8 months. A basic text module can be implemented in 2 months. Contact us to evaluate your project — we will prepare a commercial proposal considering platform and data specifics. Order turnkey development: get a consultation on architecture, timelines, and ethical aspects. We'll evaluate the project for free. Submit a pilot request — we will set up demo access within 2 weeks.

Typical mistakes when implementing AI depression detection

• Using only text without audio — loses up to 30% accuracy.
• Ignoring cultural differences — the model may misinterpret emotions.
• Lack of an ethical passport — risk of stigmatization and legal issues.
• Insufficient threshold calibration — high false positive rate.
• No clear protocol for suicidal signals.

Avoid these mistakes, and the system will be a reliable assistant for psychologists, not a source of problems.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.