Silxxor/Russian-Addressee-detector

Russian Addressee Detection Model

Fine-tuned RuBERT for detecting whether speech is directed at the system or not in Russian voice conversations.

Model Description

This model classifies whether a spoken utterance is directed at the voice assistant (addressee) or is ambient speech (talking to yourself, someone else in the room, or just mumbling). Designed for real-time voice chat applications to prevent false triggers.

Base Model: DeepPavlov/rubert-base-cased
Task: Binary classification (addressed to system / not addressed to system) Language: Russian

Performance

Metric	Score
Accuracy	95.16%
Precision	88.89%
Recall	94.12%
F1-Score	91.43%

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Silxxor/Russian-Addressee-detector")
model = AutoModelForSequenceClassification.from_pretrained("Silxxor/Russian-Addressee-detector")

text = "покажи мне погоду на завтра"
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 = not addressed to system, 1 = addressed to system
print("Addressed to system" if prediction == 1 else "Not addressed to system")

Training Data

• Source: Synthetically generated dataset based on Russian conversational patterns
• Wake words: "Люси" (Lucy), "ассистент" (assistant), "компьютер" (computer)
• Mixed dataset of direct commands, questions with wake words, and ambient speech
• Addressed utterances (containing wake words or direct commands) labeled as 1
• Non-addressed utterances (self-talk, background conversation) labeled as 0
• Balanced dataset

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 generated data, not natural speech patterns
• Optimized for specific wake words (Lucy, assistant, computer)
• May not generalize well to other assistant names or contexts
• Performance depends on ASR quality
• Cultural and contextual nuances may affect accuracy

Intended Use

Voice assistants and conversational AI systems that need to distinguish between speech directed at them versus ambient conversation in Russian.