Automatic Speech Recognition, or ASR, is the use of Machine Learning or Artificial Intelligence (AI) technology to process human speech into readable text. The field has grown exponentially over the past decade, with ASR systems popping up in popular applications we use every day such as TikTok and Instagram for real-time captions, Spotify for podcast transcriptions, Zoom for meeting transcriptions, and more.
As ASR quickly approaches human accuracy levels, there will be an explosion of applications taking advantage of ASR technology in their products to make audio and video data more accessible. Already, Speech-to-Text APIs like AssemblyAI are making ASR technology more affordable, accessible, and accurate.
This article aims to answer the question: What is ASR?, and provide a comprehensive overview of Automatic Speech Recognition technology, including:
- A History of ASR
- How ASR Works
- ASR Key Terms and Features
- Key Applications of ASR
- Challenges of ASR Today
- On the Horizon for ASR
A History of ASR
ASR as we know it extends back to 1952 when the infamous Bell Labs created “Audrey,” a digit recognizer. Audrey could only transcribe spoken numbers, but a decade later, researchers improved upon Audrey so that it could transcribe rudimentary spoken words like “hello”.
For most of the past fifteen years, ASR has been powered by classical Machine Learning technologies like Hidden Markov Models. Though once the industry standard, accuracy of these classical models had plateaued in recent years, opening the door for new approaches powered by advanced Deep Learning technology that’s also been behind the progress in other fields such as self-driving cars.
In 2014, Baidu published the infamous paper, Deep Speech: Scaling up end-to-end speech recognition. In this paper, the researchers demonstrated the strength of applying Deep Learning research to power State-of-the-Art, accurate Speech Recognition systems. The paper kicked off a renaissance in the field of ASR, popularizing the Deep Learning approach and pushing model accuracy past the plateau and closer to human level.
Not only has accuracy skyrocketed, but access to ASR technology has also improved dramatically. Ten years ago, customers would have to engage in lengthy, expensive enterprise software contracts to license ASR technology. Today, developers, startup companies, and Fortune 500s have access to State-of-the-Art ASR technology via simple APIs like AssemblyAI’s Speech-to-Text API.
Let’s look more closely at these two dominant approaches to ASR.
How ASR Works
Today, there are two main approaches to Automatic Speech Recognition: a traditional hybrid approach and an end-to-end Deep Learning approach.
Traditional Hybrid Approach
The traditional hybrid approach is the legacy approach to Speech Recognition and has dominated the field for the past fifteen years. Many companies still rely on this traditional hybrid approach simply because it’s the way it has always been done--there is more knowledge around how to build a robust model because of the extensive research and training data available, despite plateaus in accuracy.
Here’s how it works:
Traditional HMM and GMM systems
Traditional HMM (Hidden Markov Models) and GMM (Gaussian Mixture Models) require forced aligned data. Force alignment is the process of taking the text transcription of an audio speech segment and determining where in time particular words occur in the speech segment.
As you can see in the above illustration, this approach combines a lexicon model + an acoustic model + a language model to make transcription predictions.
Each step is defined in more detail below:
The lexicon model describes how words are pronounced phonetically. You usually need a custom phoneme set for each language, handcrafted by expert phoneticians.
The acoustic model (AM), models the acoustic patterns of speech. The job of the acoustic model is to predict which sound or phoneme is being spoken at each speech segment from the forced aligned data. The acoustic model is usually of an HMM or GMM variant.
The language model (LM) models the statistics of language. It learns which sequences of words are most likely to be spoken, and its job is to predict which words will follow on from the current words and with what probability.
Decoding is a process of utilizing the lexicon, acoustic, and language model to produce a transcript.
Downsides of Using the Traditional Hybrid Approach
Though still widely used, the traditional hybrid approach to Speech Recognition does have a few drawbacks. Lower accuracy, as discussed previously, is the biggest. In addition, each model must be trained independently, making them time and labor intensive. Forced aligned data is also difficult to come by and a significant amount of human labor is needed, making them less accessible. Finally, experts are needed to build a custom phonetic set in order to boost the model’s accuracy.
End-to-End Deep Learning Approach
An end-to-end Deep Learning approach is a newer way of thinking about ASR, and how we approach ASR here at AssemblyAI.
How End-to-End Deep Learning Models Work
With an end-to-end system, you can directly map a sequence of input acoustic features into a sequence of words. The data does not need to be force-aligned. Depending on the architecture, a Deep Learning system can be trained to produce accurate transcripts without a lexicon model and language model, although language models can help produce more accurate results.See: Training Deep Learning Models
CTC, LAS, and RNNT
CTC, LAS, and RNNTs are popular Speech Recognition end-to-end Deep Learning architectures. These systems can be trained to produce super accurate results without needing force aligned data, lexicon models, and language models.Learn more: Comparing End-to-End Speech Recognition Architectures
Advantages of End-to-End Deep Learning Models
End-to-end Deep Learning models are easier to train and require less human labor than a traditional approach. They are also more accurate than the traditional models being used today.
The Deep Learning research community is actively searching for ways to constantly improve these models using the latest research as well, so there’s no concern of accuracy plateaus any time soon--in fact, we’ll see Deep Learning models reach human level accuracy in the next few years.
ASR Key Terms and Features
Acoustic Model: The acoustic model takes in audio waveforms and predicts what words are present in the waveform.
Language Model: The language model can be used to help guide and correct the acoustic models predictions.
Word Error Rate: The industry standard measurement of how accurate an ASR transcription is, as compared to a human transcription.
Speaker Diarization: Answers the question, who spoke when? Also referred to as speaker labels.
Custom Vocabulary: Also referred to as Word Boost, custom vocabulary boosts accuracy for a list of specific keywords or phrases when transcribing an audio file.
Sentiment Analysis: The sentiment, typically positive, negative, or neutral, of specific speech segments in an audio or video file.
See more features specific to AssemblyAI.
Key Applications of ASR
The immense advances in the field of ASR has seen a correlation of growth in Speech-to-Text APIs. Companies are using ASR technology for Speech-to-Text applications across a diverse range of industries. Some examples include:
Telephony: Call tracking, cloud phone solutions, and contact centers need accurate transcriptions, as well as innovative analytical features like Conversational Intelligence, call analytics, speaker diarization, and more.
Video Platforms: Real-time and asynchronous video captioning are industry standard. Platforms (and video editors alike) also need content categorization and content moderation to improve accessibility and search.
Media Monitoring: Speech-to-Text APIs can help broadcast TV, podcasts, radio, and more quickly and accurately detect brand and other topic mentions for better advertising.
Virtual Meetings: Meeting platforms like Zoom, Google Meet, WebEx, and more need accurate transcriptions and the ability to analyze this content to drive key insights and action.
Choosing a Speech-to-Text API
With more APIs on the market, how do you know which Speech-to-Text API is best for your application?
Key considerations to keep in mind include:
- How accurate the API is.
- What additional features are offered.
- What kind of support you can expect.
- Pricing and documentation transparency.
- Data security.
- Company innovation.
Challenges of ASR Today
One of the main challenges of ASR today is the continual push toward human accuracy levels. While both ASR approaches--traditional hybrid and end-to-end Deep Learning--are significantly more accurate than ever before, neither can claim 100% human accuracy. This is because there is so much nuance in the way we speak, from dialects to slang to pitch. Even the best Deep Learning models can’t be trained to cover this long tail of edge cases without significant effort.
Some think they can solve this accuracy problem with custom Speech-to-Text models. However, unless you have a very specific use case, like children’s speech, custom models are actually less accurate, harder to train, and more expensive in practice than a good end-to-end Deep Learning model.
Another top concern is Speech-to-Text privacy for APIs. Too many large ASR companies use customer data to train models without explicit permission, raising serious concerns over data privacy. Continual data storage in the cloud also raises concerns over potential security breaches, especially if raw audio or video files or transcription text contains Personally Identifiable Information.
On the Horizon for ASR
As the field of ASR continues to grow, we can expect to see greater integration of Speech-to-Text technology into our everyday lives, as well as more widespread industry applications.
In regards to model building, we also expect to see a shift to a self-supervised learning system to solve some of the challenges with accuracy discussed above.
End-to-end Deep Learning models are data hungry. Our ASR model at AssemblyAI, for example, is trained on 100,000 hours of raw audio and video training data for industry-best accuracy levels. However, obtaining human transcriptions for this same training data would be almost impossible given the time constraints associated with human processing speeds.
This is where self-supervised deep learning systems can help. Essentially, this is a way to get an abundance of unlabeled data and build a foundational model on top of it. Then, since we have statistical knowledge of the data, we can fine-tune it on downstream tasks with a smaller amount of data, making it a more accessible approach to model building. This is an exciting possibility with profound implications on the field.
If this transition occurs, expect ASR models to become even more accurate and affordable, making their use and acceptance more widespread.