Lower Frame Rate Neural Network Acoustic Models

By Golan Pundak, Tara N. Sainath. 2016

144 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Speech & Audio

In Signal Processing

Citations in the corpus

Toward Domain-Invariant Speech Recognition via Large Scale Training (2018, methodology): “Input features are also subsam-pled by a factor of 3; the acoustic model operates at 33 Hz [22].”
State-of-the-Art Speech Recognition with Sequence-to-Sequence Models (2017, methodology): “Finally, we compare the proposed LAS model in E8 to a state-ofthe-art, discriminatively sequence-trained low frame rate (LFR) system [28] in terms of WER.”
Streaming End-to-end Speech Recognition for Mobile Devices (2018, methodology): “Similar to [40], at the current frame, t, these features are stacked with 3 frames to the left”
Improved training of end-to-end attention models for speech recognition (2018, background): “However, as the original input is continuous, some sort of downscaling in the time dimension works, such as concatenation in the feature dimension of consecutive time-frames [7,24,42,50].”
An Analysis of Incorporating an External Language Model into a Sequence-to-Sequence Model (2017, background): “Similar to [23, 24], at the current frame, t, these features are stacked with 3 frames to the left and downsampled to a 30ms frame rate.”
Bytes Are All You Need: End-to-end Multilingual Speech Recognition and Synthesis with Bytes (2018, background): “Similar to [29, 30], at each current frame, these features are stacked with 3 consecutive frames to the left and then down-sampled to a 30ms frame rate.”
An Attention-Based Joint Acoustic and Text on-Device End-To-End Model (2020, methodology): “These utterances are decoded by a state-of-the-art conventional ASR model [27], and thus the transcription is not ground-truth.”
Recent progresses in deep learning based acoustic models (2017, methodology): “Recently, LFMMI [80] and lower frame rate (LFR) LSTM-RNN [181] were proposed to work on the 30ms units instead of the traditional 10ms units by modeling phonemes instead of states.”
A Streaming On-Device End-To-End Model Surpassing Server-Side Conventional Model Quality and Latency (2020, methodology): “The conventional model [11] (B0) is also listed.”
Speech Recognition via Large-Scale Weak Supervision (unknown): “This happens because A very simple baseline algorithm is a method sometimes called low frame rate (Pundak and Sainath, 2016): for time i we low frame rate stack (concatenate) the acoustic feature vector f i with the prior two vectors f i − 1 and f i − 2 to make a new vector three times longer.”
A Comparison of Techniques for Language Model Integration in Encoder-Decoder Speech Recognition (2018, methodology): “We use 80-dimensional log-mel ﬁlterbank features, computed with a 25ms window and shifted every 10ms. Similarly to [29,30], at each frame t , these features are stacked with 3 frames to the left and downsam-pled to a 30ms frame rate.”
Fast, Simpler and More Accurate Hybrid ASR Systems Using Wordpieces (2020, result): “CTC [17] has became an alternative criterion to frame-level cross-entropy (CE) training or sequence-level lattice-free MMI (LF-MMI) training in recent years and has shown promising results [18–22].”
Two-Pass End-to-End Speech Recognition (2019, background): “However, this model still lags behind a large state-of-the-art conventional model in quality [2].”
Semi-supervised Training for End-to-end Models via Weak Distillation (2019, methodology): “We also present the performance of the full stack conventional model [30] we used as the teacher model for weak distillation.”
Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition (2018, background): “The remaining training features [34] from speech waveform to train multi-layer LSTM acoustic models.”
Semi-Supervised Training of Acoustic Models Using Lattice-Free MMI (2018, background): “The performance is shown to degrade with smaller datasets [4].”
End-to-End Training of Acoustic Models for Large Vocabulary Continuous Speech Recognition with TensorFlow (2017, methodology): “Front-end: The training examples are 480-dimensional frames output by the front end every 30 ms, generated by stacking four frames of 128 log mel filterbank energies extracted from a 32 ms window shifted over the audio samples in 10 ms increments [16].”
An Investigation Into On-device Personalization of End-to-end Automatic Speech Recognition Models (2019, background): “3 consecutive frames of 80-dimensional log Mel features (extracted every 10 milliseconds) are stacked together to form the 240-dimensional inputs to the network [24].”