When Forgetting Looks Like Improvement: Metric Masking in Streaming Diarizer Adaptation and the Price of Rehearsal

We present a real-time pipeline of open source, pretrained models for streaming identification of a target speaker, motivated by hearing-aid applications where latency as low as 10 ms can be perceptible. We formulate a two-step approach in which incoming audio is first segmented by speaker using low-latency streaming diarization, followed by speaker verification against a registered target speaker. To emulate conversational speech while minimizing overlap, we use the This American Life Podcast Transcripts dataset and select the host as a consistent target speaker. We benchmark offline diarization with Pyannote and LIUM using diarization error rate (DER) and select Pyannote based on baseline performance and compatibility with streaming. We then evaluate speaker verification using Pyannote and TitaNet-Large and generate ROC curves to select an operating region. We integrate Diart and tune clustering parameters to reduce DER while maintaining real-time operation. We pair Diart with Pyannote verification and evaluate system-level performance by converting predicted and ground-truth speech regions into 100 ms binary masks. Across 17 evaluation episodes, the system achieves greater than 0.90 median accuracy with high specificity (0.95-0.98) at cosine distance thresholds of 0.7-0.75, demonstrating a practical proof of concept for downstream low-latency selective amplification.

Real-time conversational agents require speaker diarization that streams and runs on a CPU. Most systems apply an utterance-level speaker encoder to short, heavily overlapping chunks, which wastes computation and leaves the model optimized for the wrong task. We instead turn a state-of-the-art speaker recognition architecture into a causal frame-level encoder that reads the stream once and emits one embedding every 80~ms from a bounded two-second window of past audio, and pair it with online clustering that gates every update on the self-similarity of the stream. Trained only by distillation from an utterance-level teacher on simulated and out-of-domain mixtures, and evaluated with one fixed set of hyperparameters, the system is the most accurate streaming diarizer on low-overlap benchmarks at sub-second latency, degrades far less than cache-based systems as the number of speakers grows, and runs five times faster than real time on a single CPU thread.

This paper describes our self-designed system for Task 1 of the MLC-SLM 2026 Challenge for multilingual two-speaker conversational speech. The system combines a modular speaker diarization front end with a challenge-adapted Qwen3-ASR-1.7B recognizer. The diarization front end performs voice activity detection, subsegment generation, CAMPPlus speaker embedding extraction, two-speaker spectral clustering, and RTTM-based audio segmentation. The resulting speaker-attributed segments are grouped by language or region and decoded by the adapted ASR model. For ASR adaptation, we first perform supervised full fine-tuning on the official training data, then apply LoRA fine-tuning with synthetic speech generated by a three-pipeline TTS-based synthetic speech augmentation framework, and finally refine the model using GRPO reinforcement learning with rewards based on WER/CER and penalties for hallucination, repetition, and length deviation. On the official development set, the full system achieves an average tcpMER of 23.70, reducing the error rate by 6.83 absolute points relative to the released Qwen-ASR-1.7B performance. On the final evaluation set, the system achieves an average tcpMER of 17.97. Ablation results show that supervised fine-tuning provides the largest gain, while synthetic-speech LoRA adaptation and reinforcement learning further improve robustness.