feanet/eou-detector-russian

Overview

This model, developed by Simplexphone, is an End-of-Utterance (EOU) detector specifically designed for Russian conversational speech in call-center environments. It is fine-tuned from the Qwen2.5-0.5B-Instruct model on a dataset of 200,000 real call-center dialogs. Unlike traditional methods, it predicts the probability of an utterance ending by detecting the <|im_end|> token at the last position, leveraging the LM vocabulary without a separate classification head.

Key Capabilities & Performance

• High Accuracy: Achieves an F1 score of 0.851 on held-out call-center data, with a False Rejection rate of 6.0% and False Alarm rate of 22.3% at an optimal threshold of 0.077.
• Low Latency: Offers fast inference, with 10 ms on H100 GPU (PyTorch) and **6 ms** with ONNX Runtime on H100 GPU for batch=1.
• Production Ready: Supports ONNX export for optimized, lower-latency deployment in production systems, including C++ integration with ONNX Runtime.
• Specialized Training: Trained using causal language modeling on the full Qwen2.5 chat template, learning to predict turn boundaries naturally without artificial labels.

Ideal Use Cases

• Voice Assistants/IVR Systems: Crucial for detecting when a caller has finished speaking before routing to ASR or NLU.
• Call-Center Analytics: Useful for segmenting transcripts by speaker turn for analysis.
• Real-time Dialog Systems: Provides a language-aware alternative to silence-based Voice Activity Detection (VAD).

Limitations

• Primarily optimized for Russian call-center speech; performance on other domains or languages is not guaranteed.
• Operates on ASR transcript text, requiring a separate VAD/ASR stage upstream.
• May exhibit unreliable scoring for very short utterances (less than 3 tokens).