Categories: Engineering

Post-processing in automatic speech recognition systems


Automatic Speech Recognition (ASR) systems provide text transcriptions. Usually, it’s a sequence of words. Cisco uses ASR systems to provide real-time closed captioning in Webex meetings. One problem that arises is that it may be difficult to read captions without punctuation and capitalization. The ability to understand the meaning of text can change based on punctuation. Imagine the following word sequence with two options for punctuation:

“thank you your donation just helped someone get a job”.

Option A: “Thank you! Your donation just helped someone get a job.”

Option B: “Thank you! Your donation just helped someone. Get a job.”

One punctuation mark makes a big difference.

We’ll walk through several considerations when building a post-processing system:

  • High-accuracy models for punctuation restoration and capitalization from raw text.
    Fast inference on interim results: to keep up with real-time captions.
  • Small resources utilization: speech recognition is computationally intensive; we don’t need our punctuation models to be computationally intensive as well.
  • Ability to process out-of-vocabulary words: sometimes, we’ll need to punctuate or capitalize words that our model hasn’t seen before.

Some classical n-gram based approaches [1] have relatively good quality. However, they come with their downsides. Although n-gram models have fast inference, even 3-gram models can take up to several gigabytes of disk space depending on the language vocabulary. Another drawback is the handling of out-of-the-vocabulary words. If a word wasn’t presented in the training data, then a model can’t process it in a regular manner, and accuracy can degrade.

Modern approaches use effective but computationally intensive techniques like the bidirectional RNN [3] or attention and transformer-based neural networks architectures [2]. These models have high accuracy [2] but may not be well suited for live streaming use cases as they require the entirety of the input sequence to run inference. For example, when you have only one new input token for a bidirectional RNN, you need to update hidden states of all tokens that model saw earlier (figure 1).

Figure 1. Computations on the backward pass of a bi-directional RNN. For every new input, all previous hidden states must be sequentially updated.


Some approaches attempt to solve punctuation and capitalization by building two different models [3][6], and others combine both into a single model since the outputs are highly correlated [4][2]. Words immediately following punctuation demonstrate this correlation well: words after periods are likely capitalized, and words after commas are likely lowercased. There are approaches that suggest an architecture with multiple outputs [4]: one per task, respectively. They show that this architecture outperforms separate punctuation and capitalization architectures.

Given the above considerations, we opted to use a single GRU-based neural network with two outputs for punctuation and capitalization.

To deal with out-of-the-vocabulary words, we use a SentencePiece-like technique [6] that splits unknown words into smaller tokens, or, in extreme cases, into characters. We describe the details and considerations below.


Intuition and experiments show that it’s essential to have future context when building a punctuation model because it’s harder to determine punctuation marks in a current position without knowing the next several words. To use information about the next tokens and not be forced to update all hidden states for all tokens in the backward direction, we decided to truncate the backward direction to a fixed window. In the forward direction, it’s just a regular RNN. In the backwards direction, we only consider a fixed window at each token, running the RNN over this window (figure 2). Using this window, we can achieve constant time inference for a new input token (we’ll need to compute one hidden state in the forward direction and n+1 in the backward direction).

Figure 2. In this example, for each current token, only the next two are considered to compute the hidden state in the backward direction.


Now, for every token, we have hidden states for forward and backward directions, respectively. Let’s call this layer TruncBiRNN or TruncBiGRU (since we use GRU). These hidden states can be computed in constant time, which does not depend on the input length. The constant time operation is essential for the model to keep up with real-time captions.


The architecture consists of embedding layer, TruncBiGRU and unidirectional GRU layer, and fully connected layer. For the output, we use two softmax layers for punctuation and capitalization, respectively (figure 3).

Figure 3. Model architecture with window size equal to two tokens for TruncBiGRU.


For every word, the model predicts its capitalization and the punctuation mark after the word. To better synchronize these two outputs and predict capitalization, we need to know embedding from the previous token, too (to restore the punctuation mark from the previous step). Together with a custom loss function (see next section), this allows us to avoid cases where a lowercase word is produced at the beginning of a sentence.

For punctuation prediction, it’s also helpful to get the capitalization prediction of the next word. That’s why we concatenate current and next embeddings.

An output layer for punctuation predicts distribution over all punctuation marks. For our model, it’s a set:

period – a period in the middle of a sentence that doesn’t necessarily imply that the next word should be capitalized (“a.m.,””D.C.,” etc)


question mark




terminal period – a period at the end of a sentence


For capitalization, we have four classes:


upper – all letters are capitalized (“IEEE,” “NASA,” etc.)


mix_case – for words like “iPhone”

leading capitalized – words that start a sentence

The additional classes, “leading capitalized” and terminal period,” may seem redundant at first glance, but they help increase the consistency of answers related to capitalization and punctuation. The “terminal period” implies that the next capitalization answer can’t be “lower,” while “leading capitalized” means that the previous punctuation mark is a “terminal period” or question mark. These classes play an important role in the loss function.

Loss Function:
We need to optimize both capitalization and punctuation. To achieve this, we use a sum of log loss function with a coefficient:

However, as stated earlier, the outputs of a neural network may not be perfectly correlated. For example, the punctuator may predict a “terminal period” for the current word, but the capitalizer doesn’t predict “leading capitalized” for the next token. This type of mistake, while rare, can be very striking. To deal with it, we use an additional penalty term in the loss function that penalizes this type of mistake:

The first term corresponds to the probability of having “leading capitalized” after non “terminal period,” and the second to the probability of not having “leading capitalized” after “terminal period.”  This penalty sums over tokens where this error occurs.

Additionally, we pass two consecutive tensors from the previous layer to softmax layers. Given that, we can efficiently reduce penalty terms.

Finally, we have the loss function:


For training, we use text transcripts from a set of internal Webex meetings and text data from Wikipedia.

First, the training data is cleaned up and split into sentences. During training, each sample is generated from consecutive sentences and is truncated to a random length from a fixed distribution. This allows the model to see cropped phrases during training which allows the model to deal with interim results during inference. Next, we train a model on approximately 300 megabytes worth of Wikipedia text, then fine-tune it on Webex meeting transcripts.

Pre-training on Wikipedia helps improve all punctuation classes but is especially useful on capitalization classes. We suspect this is due to the large number of proper nouns in the Wikipedia corpus.

We apply the same data preparation on our evaluation sets by concatenating sentences and truncating them to random lengths. This allows us to measure the accuracy for what we would likely see in interim transcript states.


Using relatively easy techniques with some customizations of the architecture, such as truncated GRU and an additional penalty in a loss function, we have built a model that can be run online. The reading experience of live captions is significantly improved with real-time punctuation marks and capitalization.


[1] A. Gravano, M. Jansche, and M. Bacchiani, “Restoring punctuation and capitalization in transcribed speech,” in ICASSP 2009, 2009, pp. 4741–4744. [2] Monica Sunkara, Srikanth Ronanki, Kalpit Dixit, Sravan Bodapati, Katrin Kirchhoff, “Robust Prediction of Punctuation and Truecasing for Medical ASR” [3] Tilk, Ottokar & Alumäe, Tanel. (2016). Bidirectional Recurrent Neural Network with Attention Mechanism for Punctuation Restoration. 3047-3051. 10.21437/Interspeech.2016-1517. [4] Vardaan Pahuja, Anirban Laha, Shachar Mirkin, Vikas Raykar, Lili Kotlerman, Guy Lev “Joint Learning of Correlated Sequence Labelling Tasks Using Bidirectional Recurrent Neural Networks” [5] Wang, Peilu & Qian, Yao & Soong, Frank & He, Lei & Zhao, Hai. (2015). Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Recurrent Neural Network. [6] Lita, Lucian & Ittycheriah, Abe & Roukos, Salim & Kambhatla, Nanda. (2003). tRuEcasIng. 10.3115/1075096.1075116. [7]

Sign up for Webex

Visit our home page or contact us directly for assistance.

Click here to learn more about the offerings from Webex and to sign up for a free account

Published by
Pavel Pekichev

Recent Posts

  • Engineering

Modern Video-Conferencing Systems: An Introduction to the Session Description Protocol

This series focuses on SDP, the Session Description Protocol, the method by which almost all…

1 day ago
  • Customer Stories
  • Hybrid Work

Cisco named a Customers’ Choice in 2024 Gartner Peer Insights™ Voice of the Customer for Meeting Solutions

We’re excited to share that Cisco was recently named a 2024 Customers’ Choice in the…

4 days ago
  • Engineering
  • Innovation

Cisco AI researchers publish a novel crowdsourced speech intelligibility test framework at ICASSP 2024

Cisco is committed to continuously advancing state-of-the-art speech enhancement AI technology, such as Background Noise…

1 week ago
  • Hybrid Work
  • Video Conferencing

Webex Video Messaging (Vidcast) | Empowering seamless workflow with asynchronous collaboration

In this hybrid work environment, we strive for work-life balance by making our day more…

1 week ago
  • Collaboration
  • Hybrid Work

Refine your Workflow with Powerful Webex Partner Integrations

Last week, Orlando was teeming with Webex partners, customers, and IT pros from across different…

2 weeks ago
  • Collaboration Devices
  • Hybrid Work
  • Innovation

More than a desk phone | The new Desk Phone 9800 Series

When I joined the calling team at Cisco - it’s hard to believe 18 years…

3 weeks ago