Improving voice trigger detection with metric learning

September 12, 2022

Voice trigger detection plays a crucial role, which enables voice assistant activation when a target user says a keyword phrase. A detector is typically trained on speech data independent of speaker information and used for the voice trigger detection task. A speaker-independent voice trigger detector, however, often performs worse on a speech from underrepresented groups.

In this blog, we discuss a novel voice trigger detector model that can increase detection accuracy by using a limited number of utterances from a target speaker. This voice trigger detection model uses an encoder-decoder architecture in contrast to traditional detectors, which conduct speaker-independent voice trigger detection, and metric learning-trained decoders anticipate a unique embedding for each syllable. By adjusting to the target speaker's speech when calculating the voice trigger score, the tailored embedding increases the precision of voice trigger detection.

How to improve Voice Trigger Detection using metic learning?

As discussed above, to improve voice trigger detection using metric learning we will be using an encoder & decoder.

A speaker-independent phoneme prediction is performed by an encoder while a decoder provides speaker-adapted voice trigger detection in the proposed MTL method for enhancing voice trigger detection using metric learning.

Down below we will understand the functions of Encoder and Decoder separately in detail.

Encoder

The encoder is made of N stacked Transformer encoder blocks with self-attention. The phoneme predictions carried out by the self-attention encoder convert the input feature sequence, represented by X, into hidden representations as:

where IN denotes a hidden representation after the n-th encoder block. In order to get logits for phoneme classes, a linear layer is applied to the final encoder output IN.

Decoder

Since the speaker information might be reduced at the top encoder layer, we utilise an intermediate representation (n<N). The trainable vectors are denoted by {qm||m = 1, ..., M}  where qm ∈ Rdx1. The encoder output and the query vectors are fed together to produce a set of decoder embedding vectors as

The output of the P-stacked Transformer decoder block is indicated by em ∈ Rdx1. The decoder output set is then modified. to create a size dM × 1 utterance-wise embedding vector. 

Then, two task-level linear layers are branched at this point: First, the embedding is covered with a linear layer to forecast a scalar logit for the keyword phrase. The second linear layer is added for speaker verification and obtaining logits. Additionally, we employ the decoder embedding and metric learning is carried out in a mini-batch.

Multi-task learning

We first introduced the MTL framework for keyword spotting, and now we propose the metric-learning loss to compare the decoder embeddings and provide a speaker-adapted speech trigger detection score. The model is trained using phonetic loss in the MTL framework at the encoder output. There are three branches available at the decoder output: speaker identification, keyword-phrase loss, and metric learning loss. 

The following can be used to formulate the training's objective function:

where the terms L(phone), L(spkr), L(phrase), and L(metric) stand for the respective phonetic loss, speaker-identification loss, keyword-phrase loss, and metric learning loss. The scaling variables for balancing the losses are α, β, and γ. 

L(phone)is a phonetic loss to compute a voice trigger detection score, L(phrase) is a loss due to cross-entropy on the scalar logits, and L(spkr) is a speaker cross-entropy loss.

The metric loss L(metric) is a cosine similarity metric. For positive pairings, referred to as utterances, the decoder embedding output is directly affected by scale and offset settings. Defines as utterances from the same speaker that contain the keyword phrase, and the negative pairings represent utterances from other speakers or utterances from the same speaker with phrase labels that are the reverse of the keyword phrase. The cosine similarity is first turned into a probability as :

where cosθij is the cosine distance between the i-th and j-th utterances' decoder embeddings. The parameters a and b stand for trainable scale and offset parameters, respectively. One may calculate Loss in metric L(metric) as:

where NP and NN represent the numbers of positive and negative pairings, and P and N represent sets of the positive and negative pairs inside a mini-batch.

Training of the model

For the purpose of training the MTL tasks, two sources of data are employed every mini-batch. The first source, which is mostly utilised for phonetic loss and keyword phrase loss, is a collection of anonymized utterances with either phoneme labels or keyword phrase labels (voice trigger data). 

The voice trigger data's non-keyword utterances are also employed as a negative class measure of learning loss. Combining an ASR dataset with the phoneme labels with a keyword spotting dataset with the keyword phrase labels will provide the dataset. The second set of data consists of speaker-labelled utterances (speaker-ID data), where each utterance is composed of a keyword phrase and a non-keyword sentence.

For each mini-batch of training, a batch sampling method is used to choose samples from each of these sets. For a batch size of 128, for instance, we choose 112 utterances from the speaker-ID data, which comprises 4 utterances from 28 different speakers, and the remaining utterances come from the voice-trigger data. In order to establish negative pairings (keyword vs. non-keyword) for the same speaker and aid in metric learning, the keyword phrase segment is also omitted at random for the utterances selected from the speaker data.

Similarity score

When making an inference, we begin by obtaining an anchor embedding, which is the average of the decoder embeddings from the speaker's prior utterances. The test utterance's decoder embedding is then calculated. 

We then determine how similar the test embedding and anchor embedding are. The speaker-adapted voice trigger score and the similarity score line up. The speaker-adapted score and a speaker-independent voice trigger score Sctc produced from the encoder output can optionally be combined. 

The speaker-adapted score is calibrated first as follows: Smetric = (Pi, anchor − C)/D, where C and D are the global parameters,  mean and standard deviation. Next, we integrate the results using a simple weighted average to combine  these two voice trigger scores:

here µ is the weighting factor.

Conclusion

In this blog, we explored a cutting-edge method for enhancing speech trigger detection by metric learning speaker information adaptation. The encoder conducts phoneme prediction for a speaker-independent voice trigger detection while the decoder predicts an utterance-wise embedding for a speaker-adapted voice trigger detection in this metric learning architecture. The decoder embedding for a test utterance and the anchor embedding for each speaker is compared to get the speaker-adapted voice trigger score. 

According to the results, metric learning surpasses the speaker-independent speech trigger detector as a baseline by 38% in terms of FRRs.

Latest Blogs
This is a decorative image for: Comparison between Cloud-Based and On Premises GPUs
October 6, 2022

Comparison between Cloud-Based and On Premises GPUs

Cloud GPUs vs On Premises GPUs

Cloud GPUs are typically more powerful than on-premises GPU instances. The cost of renting a cloud GPU is generally lower than the cost of purchasing an on-premise GPU. 

Cloud platforms offer fast access to high performance compute and deep learning algorithms, which makes it simpler to start using machine learning models and get early insights into your data. 

Cloud GPUs are better for machine learning because they have lower latency, which is important because the time it takes a neural network to learn from data affects its accuracy. Furthermore, cloud GPUs allow users to take advantage of large-scale training datasets without having to build and maintain their own infrastructure.

On Premises GPUs are better for machine learning if you need high performance or require access to cutting-edge technologies not available in the public cloud. For example, on-premises hardware can be used for deep learning applications that require high memory bandwidth and low latency.

Cloud GPUs: Cloud GPUs are remote data centers where you can rent unused GPU resources. This allows you to run your models on a massive scale, without having to install and manage a local machine learning cluster.

Lower TCO: Cloud GPUs require no upfront investment, making them ideal for companies that are looking to reduce their overall capital expenses. Furthermore, the cost of maintenance and upgrades is also low since it takes place in the cloud rather than on-premises.

Scalability & Flexibility: With cloud-based GPU resources, businesses can scale up or down as needed without any penalty. This ensures that they have the resources they need when demand spikes but also saves them money when there is little or no demand for those resources at all times.

Enhanced Capacity Planning Capabilities: Cloud GPU platforms allow businesses to better plan for future demands by providing estimates of how much processing power will be required in the next 12 months and beyond based on past data points such as workloads run and successes achieved with similar models/algorithms etc... 

Security & Compliance : Since cloud GPUs reside in a remote datacenter separate from your business' core systems, you are ensured peace of mind when it comes to security and compliance matters (eigenvector scanning / firewalls / SELinux etc...) 

Reduced Total Cost Of Ownership (TCO) over time due to pay-as-you-go pricing model which allows you only spend what you actually use vs traditional software licensing models where significant upfront investments are made.

Cloud GPUs: Cloud GPUs offer significant performance benefits over on-premises GPUs. They are accessible from anywhere, and you don't need to own or manage the hardware. This makes them a great choice for data scientists who work with multiple data sets across different platforms.

Numerous Platforms Available for Use: The wide variety of available platforms (Windows, Linux) means that you can run your models using the most popular machine learning libraries and frameworks across different platforms without having to worry about compatibility issues between them.

This is a decorative image for: Impact of the Strong Dollar: Cloud Costs Increasing, Be Indian Buy Indian
October 4, 2022

Impact of the Strong Dollar: Cloud Costs Increasing, Be Indian Buy Indian

Indian SMEs and startups are feeling the effects of the high dollar. These businesses use hyperscalers(MNC Cloud) who cannot modify their rates to account for the changing exchange rate. For certain companies, even a little shift in the currency rate may have a significant effect on their bottom line. Did you know, when the INR-USD exchange rate moved from 60 to 70 in December 2015, it had an impact of around 20% on Digital Innovation?

As the rupee is inching closer to 82 per dollar, the strong dollar has directly impacted the costs of cloud services for Indian businesses. The high cost of storage and computing power, along with bandwidth charges from overseas vendors, has led to a huge increase in the effective rate of these services. This is especially true for startups and SMEs that rely on cloud computing to store and process user data. With the strong dollar continuing to impact the cost of cloud services, it is essential for Indian companies to evaluate their options and adopt local alternatives wherever possible. This blog post will discuss how the strong dollar impacts cloud costs, as well as potential Indian alternatives you can explore in response to this global economic trend. 

What is a Strong Dollar?

A strong US dollar($) is a term used to describe a situation where a US’s currency has appreciated in value compared to other major currencies. This can be due to a variety of factors, including interest rate changes, a country’s current account deficit, and investor sentiment. When a currency appreciates, it means that it is worth more. A strong dollar makes imports more expensive, while making exports cheaper. Strong dollars have been a growing trend in the past couple of years. As the US Federal Reserve continues to hike interest rates, the dollar strengthens further. The rising value of the dollar means that the cost of cloud services, especially from hyperscalers based in the US, will rise as well. 

Increase in Cloud Costs Due to Strong Dollar

Cloud services are essential for modern businesses, as they provide easy access to software, storage, and computing resources. Cloud services are delivered over the internet and are typically charged on a per-use basis. This makes them incredibly convenient for businesses, as they can pay for only the resources they actually use. Cloud computing allows businesses to scale their resources up or down, depending on their current business needs. This makes it suitable for startups, where demand is uncertain, or large enterprises with global operations. Cloud computing is also inherently scalable and allows businesses to quickly react to changing business needs. Cloud computing is a very competitive industry and providers offer attractive prices to attract customers. However, these prices have been impacted by the strong dollar. The dollar has strengthened by 15-20% against the Indian rupee in the last few years. As a result, the costs of services such as storage and bandwidth have increased for Indian companies. Vendors charge their Indian customers in Indian rupees, taking into account the exchange rate. This has resulted in a significant rise in the costs of these services for Indian companies.

Why are Cloud Services Becoming More Expensive?

Cloud services are priced in US dollars. When the dollar is strong, the effective price of services will be higher in Indian rupees, as the cost is not re-adjusted. There are a couple of reasons for this price discrepancy. First, Indian customers will have to pay the same prices as American customers, despite a weaker Indian rupee. Second, vendors have to ensure that they make a profit.

Possible Indian Alternatives to Cloud Services

If you're looking for a cost-effective substitute for services provided by the U.S.-based suppliers, consider E2E Cloud, an Indian cloud service provider. When it comes to cloud services, E2E Cloud provides everything that startups and SMEs could possibly need.

The table below lists some of these services and compares their cost against their US equivalents. 

According to the data in the table above, Indian E2E Cloud Services are much cheaper than their American equivalents. The difference in price between some of these options is substantial. When compared to the prices charged by suppliers in the United States, E2E Cloud's bandwidth costs are surprisingly low. Although not all E2E Cloud services will be noticeably less expensive. Using Indian services, however, has an additional, crucial perk: data sovereignty.

Conclusion

The price of cloud services will rise as the US Dollar appreciates. Indian businesses will need to find ways to counteract the strong dollar's impact on their bottom lines. To do this, one must use E2E Cloud. The availability of E2E Cloud services in INR currency is a bonus on top of the already substantial cost savings. An effective protection against the negative effects of a strong dollar.

This is a decorative image for: Actions CEOs can take to get the value in Cloud Computing
September 28, 2022

Actions CEOs can take to get the value in Cloud Computing

It is not a new thing to say that a major transition is on the way. The transition in which businesses will rely heavily on cloud infrastructure rather than having their own physical IT structure. All of this is due to the cost savings and increased productivity that cloud technology brings to these businesses. Each technological advancement comes with a certain level of risk. Which must be handled carefully in order to ensure the long-term viability of the technology and the benefits it provides.

And CEOs are the primary motivators and decision-makers in any major shift or technological migration in the organization. In the twenty-first century, which is a data-driven century, it is up to the company's leader to decide what and how his/her organization will perform, overcome the risk and succeed in the coming days.

In this blog, we are going to address a few of the actions that CEOs can take to get value in cloud Computing.

  1. A Coordinated Effort

As the saying goes, the more you avoid the risk, the closer it gets. So, if CEOs and their management teams have yet to take an active part or give the necessary attention that their migration journey to the cloud requires, now is the best time to start top-team support for the cloud enablement required to expedite digital strategy, digitalization of the organization, 

The CEO's position is critical because no one else can mediate between the many stakeholders involved, including the CIO, CTO, CFO, chief human-resources officer (CHRO), chief information security officer (CISO), and business-unit leaders.

The move to cloud computing is a collective-action challenge, requiring a coordinated effort throughout an organization's leadership staff. In other words, it's a question of orchestration, and only CEOs can wield the baton. To accelerate the transition to the cloud, CEOs should ask their CIO and CTO what assistance they require to guide the business on the path.

     2. Enhancing business interactions 

To achieve the speed and agility that cloud platforms offer, regular engagement is required between IT managers and their counterparts in business units and functions, particularly those who control products and competence areas. CEOs must encourage company executives to choose qualified decision-makers to serve as product owners for each business capability.

  1. Be Agile

If your organization wants to benefit from the cloud, your IT department, if it isn't already, must become more agile. This entails more than simply transitioning development teams to agile product models. Agile IT also entails bringing agility to your IT infrastructure and operations by transitioning infrastructure and security teams from reactive, "ticket-driven" operations to proactive models in which scrum teams create application programme interfaces (APIs) that service businesses and developers can consume.

  1. Recruiting new employees 

CIOs and CTOs are currently in the lead due to their outstanding efforts in the aftermath of the epidemic. The CEOs must ensure that these executives maintain their momentum while they conduct the cloud transformation. 

Also, Cloud technology necessitates the hire of a highly skilled team of engineers, who are few in number but extremely expensive. As a result, it is envisaged that the CHRO's normal hiring procedures will need to be adjusted in order to attract the proper expertise. Company CEOs may facilitate this by appropriate involvement since this will be critical in deciding the success of the cloud transition.

  1. Model of Business Sustainability 

Funding is a critical component of shifting to the cloud. You will be creating various changes in your sector, from changing the way you now do business to utilizing new infrastructure. As a result, you'll have to spend on infrastructure, tools, and technologies. As CEO, you must develop a business strategy that ensures that every investment provides a satisfactory return on investment for your company. Then, evaluate your investments in order to optimise business development and value.

  1. Taking risks into consideration 

Risk is inherent in all aspects of corporate technology. Companies must be aware of the risks associated with cloud adoption in order to reduce security, resilience, and compliance problems. This includes, among other things, engaging in comprehensive talks about the appropriate procedures for matching risk appetite with technological environment decisions. Getting the business to take the correct risk tone will necessitate special attention from the CEO.

It's easy to allow concerns about security, resilience, and compliance to stall a cloud operation. Instead of allowing risks to derail progress, CEOs should insist on a realistic risk appetite that represents the company plan, while situating cloud computing risks within the context of current on-premises computing risks and demanding choices for risk mitigation in the cloud.

Conclusion

In conclusion, the benefits of cloud computing may be obtained through a high-level approach. A smooth collaboration between the CEO, CIO, and CTO may transform a digital transformation journey into a profitable avenue for the company.

CEOs must consider long-term cloud computing strategy and ensure that the organization is provided with the funding and resources for cloud adoption. The right communication is critical in cloud migration: employees should get these communications from C-suite executives in order to build confidence and guarantee adherence to governance requirements. Simply installing the cloud will not provide value for a company. Higher-level executives (particularly the CEO) must take the lead in the digital transformation path.

Build on the most powerful infrastructure cloud

A vector illustration of a tech city using latest cloud technologies & infrastructure