arXiv Computation and Language

Benchmarking Speaker Embedders in the Sub-Second Regime for Real-Time Voice AI Agents

The paper evaluates several open‑source speaker embedding models—TitaNet‑Large, TitaNet‑Small, ECAPA‑TDNN, CAM++, and SpeakerNet—under a strict sub‑second latency constraint for real‑time voice AI agents. Using a controlled benchmark that injects a second speaker into real call recordings and a weighted human‑annotated production set, the authors find that (1) denoising before embedding generally worsens performance, (2) truncating audio below the full voice‑activity segment degrades all models, and (3) about half of the TitaNet embedding dimensions are redundant, encoding demographic attributes that add little discriminative value. They recommend running verification on raw, full voice‑activity audio, setting thresholds to favor precision, and using truncated embeddings to reduce similarity‑search cost.

arXiv AI
Aug 18

Adding Voice Cloning to Text-to-Audio-Video Models with a Single Zero-Initialised Layer

arXiv:2608. 15690v1 Announce Type: cross Abstract: Text-to-audio-video (T2AV) generation models produce a video and its soundtrack from a textual description, but offer no control over whose voice speaks in the output.

By Ivan Mikheev, Viacheslav Vasilev, Anna Dmitrienko, Alexey Letunovskiy, Ivan Kirillov, Kirill Chernyshev, Denis Dimitrov
arXiv AI
Jul 17

RW-Voice-EQ Bench: A Real World Benchmark for Evaluating Voice AI Systems

arXiv:2607. 14846v1 Announce Type: cross Abstract: Current voice AI benchmarks typically evaluate isolated capabilities such as speech intelligibility, word error rate, or text-based dialogue quality, but they rarely test whether systems harness the acoustic information that distinguishes spoken language from its textual representation.

By David Ayllon, Alice Baird, Jeffrey Brooks, Franc Camps-Febrer, Jakub Piotr C{\l}apa, Theo Lebryk, Jens Madsen, Olya Ossipova, Sharath Rao, Hoon Shin, Tigran Soghbatyan, Georg Streich, Rashish Tandon, Panagiotis Tzirakis
arXiv Computation and Language
Sep 28

Inference-Time Target Speaker Unlearning in LLM-Based Automatic Speech Recognition

The paper introduces a new target‑speaker unlearning task for automatic speech recognition (TSU‑ASR) that allows certain speakers to opt out of transcription while still indicating their presence. A lightweight Enrollment‑Conditioned Gating (ECG) module is added to a frozen dual‑stream speech LLM, enabling dynamic unlearning of new opt‑out speakers during inference. Experiments on AMI and AliMeeting datasets show significant drops in transcription accuracy for opt‑out speakers while preserving performance for retained speakers.

By Bo Su, Yueru Yan, Thai Le
arXiv AI
Jul 14

Breaking the Quality--Intelligibility Trade-off in Streaming Target Speaker Extraction via Deep-Feature-Anchored Preference Optimization

arXiv:2607. 10191v1 Announce Type: cross Abstract: Generative streaming models for Target Speaker Extraction (TSE) commonly exhibit a quality--intelligibility trade-off, wherein naive optimization for perceptual audio quality tends to degrade speech intelligibility, and conversely.

By Shuhai Peng, Jinjiang Liu, Hui Lu, Liyang Chen, Guiping Zhong, Jiakui Li, Shiyin Kang, Zhiyong Wu
arXiv Machine Learning
Oct 5

FASTDIAR: Frame-level speaker encoder for Streaming Diarization

FASTDIAR is a frame‑level speaker encoder designed for real‑time streaming diarization on a single CPU thread. It converts a state‑of‑the‑art speaker recognition architecture into a causal model that emits an embedding every 80 ms from a two‑second window, and pairs it with online clustering that updates only when the stream’s self‑similarity changes. Trained via distillation from an utterance‑level teacher, it achieves the highest accuracy on low‑overlap benchmarks, degrades less with more speakers, and runs five times faster than real time.

By Nikita Torgashov, Okan K\"op\"ukl\"u
arXiv AI
Sep 10

X2Streaming-ASR: wait when uncertain, emit when ready for streaming ASR

X2Streaming-ASR introduces a method for streaming automatic speech recognition that separates the decision of when to commit a transcript from what to commit. The approach uses a three‑stage training process: first establishing streaming capability, then warm‑starting a commit policy with automatically probed trajectories, and finally refining the policy with character‑level, segment‑assigned group‑relative rewards for accuracy and latency. On AISHELL‑1/2/3 and WenetSpeech datasets, the system achieves mean character‑level commit latencies of 27–84 ms, far lower than baseline systems, while also attaining the best streaming character error rates on AISHELL‑1 and AISHELL‑3.

By Zhiwei Lin, Kaiqi Fu, Rime Wen, Zehan Liu, Shawn Qin, Roy Gan, Hao Wang, Qian Wang
arXiv AI
Sep 7

X-VC: Zero-shot Streaming Voice Conversion in Codec Space

X-VC is a zero‑shot streaming voice conversion system that performs one‑step conversion directly in the latent space of a pretrained neural codec. It employs a dual‑conditioning acoustic converter that jointly models source codec latents and target acoustic conditions, while using adaptive normalization to inject utterance‑level speaker information. The model is trained with generated paired data and a role‑assignment strategy, and uses a chunkwise inference scheme with overlap smoothing to achieve low‑latency streaming inference, achieving superior WER, speaker similarity, and real‑time factor on the Seed‑TTS‑Eval benchmark.

By Qixi Zheng, Yuxiang Zhao, Tianrui Wang, Wenxi Chen, Kele Xu, Yikang Li, Qinyuan Cheng, Xipeng Qiu, Kai Yu, Xie Chen