Reed: An Approach Towards Quickly Bootstrapping Multilingual Acoustic Models

Sen, Bipasha; Agarwal, Aditya; Ganesh, Mirishkar Sai; Vuppala, Anil Kumar

doi:10.1109/slt48900.2021.9383457

Cited by 2 publications

(4 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…• Other SOTA Systems. Given that the MSR challenge was launched in 2018, there are some follow-up works (Sailor and Hain 2020;Sen et al 2021Sen et al , 2020 which have reported results on this dataset. We report the numbers as it is from these works.…”

Section: Comparison With Existing Baselinesmentioning

confidence: 93%

“…In this section, we focus on different choices for fine-tuning and decoding. Choice of Language Model As mentioned earlier, language (Sailor and Hain 2020) 18.4 16.3 18.6 Reed (Sen et al 2021) 16.1 19.9 20.2 CNN + Context temporal features (Sen et al 2020) 18 language model: (i) a LM trained using only the transcripts for the training and validation data and (ii) a LM trained on much larger data from IndicCorp in addition to the transcripts for the training and validation data. Comparing rows 5 and 6 of Table 2 we observe that integrating a LM trained on larger generic data outperforms a LM trained on smaller task specific data.…”

Section: Ablation Studies On Fine-tuning and Decodingmentioning

confidence: 99%

See 1 more Smart Citation

Towards Building ASR Systems for the Next Billion Users

Javed

Doddapaneni

Raman³

et al. 2022

AAAI

View full text Add to dashboard Cite

Recent methods in speech and language technology pretrain very large models which are fine-tuned for specific tasks. However, the benefits of such large models are often limited to a few resource rich languages of the world. In this work, we make multiple contributions towards building ASR systems for low resource languages from the Indian subcontinent. First, we curate 17,000 hours of raw speech data for 40 Indian languages from a wide variety of domains including education, news, technology, and finance. Second, using this raw speech data we pretrain several variants of wav2vec style models for 40 Indian languages. Third, we analyze the pretrained models to find key features: codebook vectors of similar sounding phonemes are shared across languages, representations across layers are discriminative of the language family, and attention heads often pay attention within small local windows. Fourth, we fine-tune this model for downstream ASR for 9 languages and obtain state-of-the-art results on 3 public datasets, including on very low-resource languages such as Sinhala and Nepali. Our work establishes that multilingual pretraining is an effective strategy for building ASR systems for the linguistically diverse speakers of the Indian subcontinent.

show abstract

Section: Comparison With Existing Baselinesmentioning

confidence: 93%

Section: Ablation Studies On Fine-tuning and Decodingmentioning

confidence: 99%

Towards Building ASR Systems for the Next Billion Users

Javed

Doddapaneni

Raman³

et al. 2022

AAAI

View full text Add to dashboard Cite

show abstract

“…• Other SOTA systems. Given that the MSR challenge was launched in 2018, there are some followup works (Sailor and Hain, 2020;Sen et al, 2021Sen et al, , 2020 which have reported results on this dataset. We report the numbers as it is from these works.…”

Section: Comparison With Existing Baselinesmentioning

confidence: 95%

“…This indicates that developing ASR systems for Indic languages with their richer phoneme/character sets and vocabulary is more challenging than English. Jilebi (Pulugundla et al, 2018) 14.0 13.9 14.7 Cogknit (Fathima et al, 2018) 17.7 16.0 17.1 CSALT-LEAP (Srivastava et al, 2018) -16.3 17.6 ISI-Billa (Billa, 2018) 19.3 19.6 20.9 MTL-SOL (Sailor and Hain, 2020) 18.4 16.3 18.6 Reed (Sen et al, 2021) 16.1 19.9 20.2 CNN + Context temporal features (Sen et al, 2020) 18…”

Section: Ablation Studies On Fine-tuning and Decodingmentioning

confidence: 99%

Towards Building ASR Systems for the Next Billion Users

Javed¹,

Doddapaneni²,

Raman³

et al. 2021

Preprint

View full text Add to dashboard Cite

Recent methods in speech and language technology pretrain very LARGE models which are fine-tuned for specific tasks. However, the benefits of such LARGE models are often limited to a few resource rich languages of the world. In this work, we make multiple contributions towards building ASR systems for low resource languages from the Indian subcontinent. First, we curate 17,000 hours of raw speech data for 40 Indian languages from a wide variety of domains including education, news, technology, and finance. Second, using this raw speech data we pretrain several variants of wav2vec style models for 40 Indian languages. Third, we analyze the pretrained models to find key features: codebook vectors of similar sounding phonemes are shared across languages, representations across layers are discriminative of the language family, and attention heads often pay attention within small local windows. Fourth, we fine-tune this model for downstream ASR for 9 languages and obtain state-of-the-art results on 3 public datasets, including on very low-resource languages such as Sinhala and Nepali. Our work establishes that multilingual pretraining is an effective strategy for building ASR systems for the linguistically diverse speakers of the Indian subcontinent.

show abstract

Reed: An Approach Towards Quickly Bootstrapping Multilingual Acoustic Models

Cited by 2 publications

References 22 publications

Towards Building ASR Systems for the Next Billion Users

Towards Building ASR Systems for the Next Billion Users

Towards Building ASR Systems for the Next Billion Users

Contact Info

Product

Resources

About