Silxxor/russian-turn-detector

Russian Turn Detection Model

Fine-tuned RuBERT for detecting complete vs incomplete utterances in real-time Russian voice conversations.

Model Description

This model classifies whether a spoken utterance is complete (ready for response) or incomplete (speaker still talking). Trained on phone call transcripts for real-time voice chat applications.

Base Model: DeepPavlov/rubert-base-cased
Task: Binary classification (complete/incomplete turn detection)
Language: Russian

Performance

Metric	Score
Accuracy	90.00%
Precision	91.33%
Recall	100.00%
F1-Score	90.91%

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Silxxor/russian-turn-detector")
model = AutoModelForSequenceClassification.from_pretrained("Silxxor/russian-turn-detector")

text = "да я понимаю что"
inputs = tokenizer(text, return_tensors="pt", max_length=64, truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=-1).item()

# 0 = incomplete turn, 1 = complete turn
print("Complete" if prediction == 1 else "Incomplete")

Training Data

• Source: ASR transcripts from Russian phone calls
• Sentence length: 3-15 words
• Complete sentences labeled as 1
• Incomplete sentences (20-60% truncated from end) labeled as 0
• Balanced dataset

Preprocessing

• Removed common Russian fillers: эээ, ммм, ааа, ну, вот, типа, короче
• Lowercased text
• Normalized whitespace
• Max sequence length: 64 tokens

Training Details

• Epochs: 3
• Batch size: 16
• Learning rate: 2e-5
• Weight decay: 0.01
• Optimizer: AdamW
• Best model selection: F1 score

Limitations

• Trained on synthetically truncated sentences, not natural speech interruptions
• Optimized for phone call scenarios with ASR artifacts
• May not generalize to other Russian speech domains
• Performance may degrade on very short (<3 words) or long (>15 words) utterances

Intended Use

Real-time voice assistants and conversational AI systems that need to determine when a speaker has finished their turn in Russian.