Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #8351

Zakaria Aldeneh

Related event timeline, sources and context from the news index.

EVENT TIMELINE

3

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Code-switching detection from speech using language-selective dimensions in Whisper-small

The authors identify language-selective hidden-state dimensions in pretrained speech encoders and aggregate their frame-level activations to detect code-switching (CS) directly from audio. Using a small tuning set, the Whisper-small encoder achieved a mean F1 of 0.798 across 14 language pairs on CS-FLEURS, outperforming a CNN (0.776) and a HuBERT linear baseline (0.751).

6.0

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Language discrimination during HuBERT pretraining reduces the bilingual multilingual gap

Authors report that strengthening language discrimination in a controlled English–French HuBERT pretraining setup—via an auxiliary language classifier or per-language k-means targets—reduces the multilingual gap on phonetic, lexical, and prosodic measures while retaining cross-language sharing. Key results: phone-ABX fell from 11.6% to 10.4% (monolingual: 10.8%), sWUGGY rose from 52.1% to 56.7% (monolingual: 58.5%), and ProsAudit (lexical) rose from 68.9% to 72.9%; largest gains occur when discrimination is introduced in the first HuBERT iteration.

7.0

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Semi-supervised federated ASR: online pseudo-labels with server update stabilization

The paper studies semi-supervised federated learning for automatic speech recognition and shows that two coupled design axes—the teacher that generates pseudo-labels (online per-client vs. broadcast global) and a server-side labeled-data 'anchor' that continues training between rounds—determine stability and final performance. With the recommended stabilization (interleaved server training and tuned augmentation/batch settings), their recipe outperforms the strongest prior method on 9 of 11 test pairs, reducing the gap to fully supervised FL by 20.8% on average in-domain and 10.0% cross-domain.

7.0