IP Library › Granted Patent US 12,725,608
Granted Patent B2
US 12,725,608 · App. 18/659,514 · Granted Sep 1, 2026

Automatic speech recognition with multilingual scalability and low-resource adaptation

Inventors: Yang Zhang (Cambridge, MA); Kaizhi Qian (Champaign, IL); Chuang Gan (Cambridge, MA); Zhenfang Chen (Cambridge, MA)
Assignee: INTERNATIONAL BUSINESS MACHINES CORPORATION
G10L15/065G10L15/063G10L15/16G10L15/005G10L2015/0635
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,725,608
App. No.
18/659,514
Granted
Sep 1, 2026
Kind
B2
Abstract

Systems and techniques that facilitate multilingual ASR machine learning models are provided. For example, one or more embodiments described herein can comprise a system, which can comprise a memory that can store computer executable components. The system can also comprise a processor, operably coupled to the memory that can execute the computer executable components stored in memory. The computer executable components can comprise an automatic speech recognition machine learning model that learns shared weights across one or more languages in a set of target languages; and a training component that trains the automatic speech recognition model.

Claims (53)

1 . A system comprising:

a memory that stores computer executable components;

a processor that executes the computer executable components stored in the memory,

wherein the computer executable components comprise:

an automatic speech recognition machine learning model that learns shared weights across one or more languages in a set of target languages, the automatic speech recognition machine learning model comprising a projection layer having pretrained weights inherited from a previously trained model; and

a training component that trains the automatic speech recognition machine learning model, wherein the training comprises:

learning, using the projection layer, a mapping matrix having a plurality of rows corresponding respectively to languages of the set of target languages and columns corresponding to individual pretrained weights of the projection layer, the mapping matrix comprising rows of specialist scores for the one or more languages in the set of target languages;

generating a set of binary masks for the one or more languages in the set of target languages by thresholding the specialist scores of the mapping matrix to identify a subset of the pretrained weights associated with each language; and

adapting the automatic speech recognition machine learning model to the one or more languages of the set of target languages by selectively activating the subset of pretrained weights of the projection layer according to the generated binary masks while retaining remaining pretrained weights as shared weights across the set of target languages.

2 . The system of claim 1 , wherein the projection layer inherits the one or more pretrained weights from at least one of a QKV (query key value layer) layer or one or more previous projection layers of the automatic speech recognition machine learning model.

3 . The system of claim 1 , wherein the computer executable components further comprise:

a tuning component that tunes the automatic speech recognition machine learning model with an additional language added to the set of target languages.

4 . The system of claim 3 , wherein the tuning comprises:

locking parameters of the projection layer;

expanding the mapping matrix with an additional row for the additional language;

leaning specialist scores for the additional row of the mapping matrix;

generating a binary mask for the additional language; and

adapting the automatic speech recognition machine learning model to the additional language by applying the binary mask to the one or more pretrained weights of the projection layer.

5 . The system of claim 1 , wherein the specialist scores comprise a shared shape with the one or more pretrained weights.

6 . The system of claim 1 , wherein the automatic speech recognition machine learning model comprises a transformer model.

7 . The system of claim 1 , wherein the shared weights represent phonetic and syntactic similarities between languages of the set of target languages.

8 . A computer-implemented method comprising:

training, by a system operatively coupled to a processor, an automatic speech recognition machine learning model to learn shared weights across one or more languages in a set of target languages, the automatic speech recognition machine learning model comprising a projection layer having pretrained weights inherited from a previously trained speech recognition model, wherein the training comprises:

learning, by the system, using the projection layer of the automatic speech recognition machine learning model, a mapping matrix having a plurality of rows corresponding respectively to languages of the set of target languages and columns corresponding to projection weights of the projection layer, the mapping matrix comprising rows of specialist scores for the one or more languages in the set of target languages;

generating, by the system, a set of binary masks for the one or more languages in the set of target languages based on the specialist scores of the mapping matrix, each binary mask identifying a subset of projection weights associated with a respective language; and

adapting, by the system, the automatic speech recognition machine learning model to the one or more languages of the set of target languages by applying the binary mask corresponding to a selected language to modulate outputs of the projection layer using the subset of projection weights identified by the binary mask while maintaining remaining projection weights as shared weights across the set of target languages.

9 . The computer-implemented method of claim 8 , wherein the projection layer inherits the one or more pretrained weights from at least one of a QKV (query key value layer) layer or one or more previous projection layers of the automatic speech recognition machine learning model.

10 . The computer-implemented method of claim 8 , further comprising tuning, by the system, the automatic speech recognition machine learning model with an additional language added to the set of target languages.

11 . The computer-implemented method of claim 10 , wherein the tuning comprises:

locking, by the system, parameters of the projection layer;

expanding, by the system, the mapping matrix with an additional row for the additional language;

leaning, by the system, specialist scores for the additional row of the mapping matrix;

generating, by the system, a binary mask for the additional language; and

adapting, by the system, the automatic speech recognition machine learning model to the additional language by applying the binary mask to the one or more pretrained weights of the projection layer.

12 . The computer-implemented method of claim 8 , wherein the specialist scores comprise a shared shape with the one or more pretrained weights.

13 . The computer-implemented method of claim 8 , wherein the automatic speech recognition machine learning model comprises a transformer model.

14 . The computer-implemented method of claim 8 , wherein the shared weights represent phonetic and syntactic similarities between languages of the set of target languages.

15 . A computer program product, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

train, by the processor, an automatic speech recognition machine learning model to learn shared weights across one or more languages in a set of target languages, the automatic speech recognition machine learning model comprising a projection layer having pretrained weights inherited from a previously trained speech recognition model, wherein the training comprises:

learning, by the processor, using the projection layer of the automatic speech recognition machine learning model, a mapping matrix having a plurality of rows corresponding respectively to languages of the set of target languages and columns corresponding to projection weights of the projection layer, the mapping matrix comprising rows of specialist scores for the one or more languages in the set of target languages;

generating, by the processor, a set of binary masks for the one or more languages in the set of target languages based on the specialist scores of the mapping matrix, each binary mask representing a language-specific selection vector corresponding to the projection weights of the projection layer; and

adapting, by the processor, the automatic speech recognition machine learning model to the one or more languages of the set of target languages by selectively applying, by the program instructions, the binary masks to the pretrained weights of the projection layer to enable language-specific specialization of the projection layer while retaining remaining pretrained weights as shared parameters across the set of target languages.

16 . The computer program product of claim 15 , wherein the projection layer inherits the one or more pretrained weights from at least one of a QKV (query key value layer) layer or one or more previous projection layers of the automatic speech recognition machine learning model.

17 . The computer program product of claim 15 , wherein the program instructions are further executable to cause the processor to:

tune, by the processor, the automatic speech recognition machine learning model with an additional language added to the set of target languages.

18 . The computer program product of claim 17 , wherein the tuning comprises:

locking, by the processor, parameters of the projection layer;

expanding, by the processor, the mapping matrix with an additional row for the additional language;

leaning, by the processor, specialist scores for the additional row of the mapping matrix;

generating, by the processor, a binary mask for the additional language; and

adapting, by the processor, the automatic speech recognition machine learning model to the additional language by applying the binary mask to the one or more pretrained weights of the projection layer.

19 . The computer program product of claim 15 , wherein the specialist scores comprise a shared shape with the one or more pretrained weights.

20 . The computer program product of claim 15 , wherein the shared weights represent phonetic and syntactic similarities between languages of the set of target languages.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded May 9, 2024
From: ZHANG, YANG; QIAN, KAIZHI; GAN, CHUANG; CHEN, ZHENFANG
To: INTERNATIONAL BUSINESS MACHINES CORPORATION
Reel/Frame 067363/0244 →
Continuity (1)
Related Publication 20250349284A1 · Nov 13, 2025
References Cited (26)
US 12354002B1 · Weber · 2025 [cited by examiner]
US 20180137109A1 · Mangoubi et al. · 2018 [cited by applicant]
US 20210294970A1 · Bender · 2021 [cited by examiner]
US 20220309340A1 · Leal · 2022 [cited by examiner]
US 20220310069A1 · Ren et al. · 2022 [cited by applicant]
US 20230214605A1 · Bérard · 2023 [cited by examiner]
US 20240005912A1 · Agrawal · 2024 [cited by examiner]
US 20250078815A1 · Ding · 2025 [cited by examiner]
US 20250095652A1 · Chen · 2025 [cited by examiner]
US 20250149026A1 · Wang · 2025 [cited by examiner]
US 20250246181A1 · Ngai · 2025 [cited by examiner]
US 20250272541A1 · Yan · 2025 [cited by examiner]
US 20250279089A1 · Liu · 2025 [cited by examiner]
Kulkarni, Atharva, et al. “Adapting the adapters for code-switching in multilingual ASR.” arXiv preprint arXiv:2310.07423, Oct. 2023, pp. 1-5. (Year: 2023). [cited by examiner]
Lee, et al. “Embedding articulatory constraints for low-resource speech recognition based on large pre-trained model.” Proc. Interspeech. Aug. 2023, pp. 1394-1398. (Year: 2023). [cited by examiner]
Lu, Yizhou, et al. “Language adaptive cross-lingual speech representation learning with sparse sharing sub-networks.” ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). I… [cited by examiner]
Qin, Siqing, et al. “Improving low-resource Tibetan end-to-end ASR by multilingual and multilevel unit modeling.” EURASIP Journal on Audio, Speech, and Music Processing Jan. 2022, Jan. 2022, pp. 1-10. (Year: 2022). [cited by examiner]
Wang, Pu. “Bottleneck low-rank transformers for low-resource spoken language understanding.” arXiv preprint arXiv:2206.14318, Jun. 2022, pp. 1-5. (Year: 2022). [cited by examiner]
Zhu, Yun, et al. “Multilingual Speech Recognition with Self-Attention Structured Parameterization.” Interspeech. Oct. 2020, pp. 4741-4745. (Year: 2020). [cited by examiner]
Fu, Yonggan, et al. “Losses can be blessings: Routing self-supervised speech representations towards efficient multilingual and multitask speech processing.” Advances in Neural Information Processing Systems 35. Dec. 20… [cited by examiner]
Xie, et al. “Dynamic asr pathways: An adaptive masking approach towards efficient pruning of a multilingual asr model.” ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).… [cited by examiner]
Yang, Mu, et al. “Learning asr pathways: A sparse multilingual asr model.” ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Sep. 2023, pp. 1-5. (Year: 2023). [cited by examiner]
Egorova, et al., “Manual and Semi-Automatic Approaches to Building a Multilingual Phoneme Set,” Published in: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, Date of Conference: May 26-31,… [cited by applicant]
Wang, et al., “MiniLMv2: Multi-Head Self-Attention Relation Distillation for Compressing Pretrained Transformers,” arXiv:2012.15828v2 [cs.CL] Jun. 27, 2021. [cited by applicant]
Andersen, et al., “On the Use of Data-Driven Clustering Technique for Identification of Poly- and Mono-phonemes for four European Languages,” Published in: Proceedings of ICASSP '94. IEEE International Conference on Aco… [cited by applicant]
Yu, et al., “Master-ASR: Achieving Multilingual Scalability and Low-Resource Adaptation in ASR with Modular Learning,” arXiv:2306.15686v1 [eess.AS] Jun. 23, 2023. [cited by applicant]