As the founder and Chief Executive Officer of Gladia. Prior to this role, he held the position of Group Vice President of Knowledge, Artificial Intelligence, and Quantum Computing at OVHcloud, a leading European cloud provider. He holds Master’s Diplomas in Symbolic Artificial Intelligence from both the University of Québec in Canada and Arts et Métiers ParisTech in France. Throughout his distinguished career, he has occupied key roles across diverse sectors, including financial data analytics, machine learning applications for real-time digital marketing, and the development of speech AI application programming interfaces.
Provides exceptional audio transcription and cutting-edge AI solutions for effortless integration across diverse industries, languages, and professional settings. By leveraging cutting-edge Automatic Speech Recognition (ASR) and generative artificial intelligence models, the technology provides seamless, real-time speech and language processing with unparalleled accuracy. Gladia’s platform enables real-time extraction of valuable insights and metadata from calls and conferences, thereby facilitating key enterprise applications such as sales support and automated customer assistance.
Initially, my intention was to create Gladia, an AI company that would democratize access to sophisticated knowledge. As our investigation progressed, a stark reality emerged: voice expertise, though initially underestimated, had become the most vulnerable yet vital area requiring urgent focus.
Voices play a vital role in our daily lives, with the majority of our communication taking place through spoken language. However, the instruments available for builders to work with voice data have fallen short in terms of speed, accuracy, and value – especially across languages.
I aimed to simplify the intricacies of voice expertise, breaking them down into a cohesive, eco-friendly, high-impact, and accessible framework. Without worrying about the intricacies of AI architectures or the subtleties of context sizes in speech recognition, builders should focus on crafting structures that are truly remarkable. I aimed to design a high-end speech-to-text API that effortlessly integrated with any hardware or software architecture – a genuine out-of-the-box solution.
In speech recognition, two critical metrics – pace and accuracy – are designed to be inversely correlated, with faster recognition typically sacrificing some level of precision and vice versa. Since pursuing one goal may come at the expense of another, it is crucial to strike a balance between them. The main cause of the fee dispute is largely a result of the supplier’s dilemma between prioritizing speed or quality.
During the development of Gladia, our goal was to strike a harmonious balance between scalability and ease of use, ensuring that the platform remains accessible to startups and small-to-medium-sized enterprises alike. As we delved deeper, we discovered that fundamental automatic speech recognition (ASR) models, such as OpenAI’s Whisper, which we worked with extensively, exhibit biases, heavily skewed towards English due to their training data, leaving many languages underrepresented and struggling for accurate transcription.
To overcome the speed-accuracy tradeoff, we needed to optimize and tailor our core processes for a multicultural team to build a global API that enables seamless business operations across languages.
Our revolutionary Gladia Actual Time engine boasts a market-defining latency of just 300 milliseconds. By leveraging audio intelligence capabilities, such as named entity recognition and sentiment analysis, this technology can effectively extract valuable insights from names or assemblies.
Our analysis indicates that only a handful of competitors possess the capability to deliver high-quality transcriptions and insights with an extremely low latency threshold, under one second for end-to-end processing, while also supporting languages beyond English. Our language support currently covers more than 100 languages across the globe.
To further enhance the versatility of our product, we have prioritized ensuring its compatibility with various stack configurations. Our API seamlessly integrates with a range of current technology stacks and telephony protocols, including SIP, VoIP, FreeSwitch, and Asterisk. The company’s telephony protocols have achieved a high level of sophistication, making them an attractive integration option for the market. This product has significant value-added potential in its product suite.
When a mannequin is lacking in data or insufficiently informed about a topic, hallucinations are likely to occur. Although fashion models can generate outputs tailored to a specific request, they are limited by the data available at the time of training and may not incorporate up-to-date information. The mannequin will generate plausible yet unreliable responses by completing gaps with information that appears authentic but lacks factual accuracy.
While hallucinations were initially detected within Large Language Models (LLMs), they also manifest in speech recognition models like Whisper ASR, a leading model developed by OpenAI. Like LLMs, Whisper’s hallucinations arise from its similar architecture, making it an issue that concerns generative models capable of predicting subsequent words based on overall context. They effectively manufacture the output from scratch. This approach will be juxtaposed against more traditional, acoustic-centric automatic speech recognition (ASR) frameworks that rely on a more rigid mapping of input sounds to output phonemes.
As a result, unwarranted phrases may be uncovered in a transcript, presenting a significant issue, especially in highly regulated domains such as medicine, where inaccuracies of this nature can have severe and far-reaching consequences.
Strategies for handling and detecting hallucinations abound. A common approach is to employ a retrieval-augmented generation (RAG) framework, combining the model’s generative prowess with a retrieval mechanism to verify facts and enhance accuracy. The chain-of-thought method involves guiding the mannequin through a series of predetermined steps or checkpoints to maintain its logical progression.
A novel approach to identifying hallucinations involves deploying algorithms that scrutinize the veracity of the model’s responses during training. Benchmarks exist specifically to assess hallucinations, involving a comparison of candidate responses generated by the model with those actually found in the input data, and determining the most accurate one.
At Gladia, we have rigorously tested a range of innovative approaches in the development of Whisper-Zero, our proprietary Automatic Speech Recognition system that successfully minimizes hallucinations. Confirmation is given that high-quality leads result in efficient asynchronous transcription, with efforts currently focused on refining this process to achieve a consistent information integrity of 99.9%, effective in real-time.
Language detection in Automatic Speech Recognition (ASR) is a particularly sophisticated task requiring advanced computational powers and nuanced algorithms to accurately identify the spoken language. Each individual boasts a unique vocal fingerprint, commonly referred to as vocalesthesia. Machine learning algorithms can perform classifications by analysing the vocal spectrum, leveraging Mel Frequency Cepstral Coefficients (MFCCs) to extract key frequency characteristics and inform their decision-making processes.
MFCCs are a technique inspired by the human auditory system’s natural processing of sound features. The field is rooted in psychoacoustics, which delves into the intricacies of human perception and comprehension of sound. The software highlights reduced frequencies by applying techniques such as normalized Fourier decomposition, converting audio signals into a frequency spectrum.
Despite its effectiveness, this method has a significant limitation: it relies solely on acoustic principles. If you converse English with a strong accent, the system may struggle to comprehend the content, instead relying heavily on prosodic cues such as rhythm, stress, and intonation.
At this location, customers can access Gladia’s progressive resolution services. By combining psychoacoustic approaches with content-aware insights, we have successfully crafted a cutting-edge methodology for real-time language recognition.
Our system doesn’t just listen to how you communicate, but also comprehends the meaning behind your words. This twin approach enables environmentally conscious code-switching, effectively preventing robust accents from being misinterpreted or misunderstood.
By seamlessly incorporating code-switching—a distinct hallmark of our approach—into multilingual interactions, we effectively bridge linguistic and cultural divides, fostering seamless communication. The audio system’s capacity to seamlessly switch between languages, even mid-conversation or sentence, underscores the paramount importance of precise transcription capabilities by the mannequin in real-time.
The Gladia API uniquely excels at handling code-switching among numerous language pairs, boasting exceptional accuracy, even in noisy environments where transcription quality is prone to decline.
To achieve latency below 300 milliseconds while maintaining high accuracy, a comprehensive approach is necessary, combining expertise in hardware, optimized algorithms, and innovative architecture.
Unlike conventional computing, actual-time AI is intricately tied to the capabilities and efficiency of Graphics Processing Units (GPGPUs). I’ve spent nearly a decade refining this property, leading the AI division at OVHCloud, Europe’s top cloud provider, and gained insight into the delicate balance between hardware power, cost, and algorithmic harmonization.
Time-efficient AI performance arises when algorithms are precisely calibrated to harmonize with hardware capabilities, ensuring each computation optimizes processing speed while minimizing latency.
However, it’s not just the AI and hardware that drive innovation. The system’s architecture plays a significant role in performance, with the community aspect having a tangible impact on latency. With decades of expertise in low-latency community architecture under his belt, courtesy of his tenure at pioneering IoT company Sigfox, our CTO has skillfully fine-tuned our community infrastructure to trim valuable milliseconds from the overall experience.
It’s a harmonious blend of judicious hardware choices, finely tuned algorithms, and collaborative design that enables us to consistently achieve sub-300ms latency without sacrificing accuracy.
Artificially intelligent conversational systems like ASR unlock a multitude of possibilities across various industries, delighting in the proliferation of innovative companies that have emerged over the past two years, utilizing Large Language Models (LLMs) and our Application Programming Interface (API) to develop groundbreaking, competitive products. Listed here are some examples:
- Purchasers are increasingly seeking to equip professionals with portable devices that enable swift recording and organization of information gathered at critical events such as work meetings, student lectures, or medical consultations? Our advanced speaker diarization technology enables precise identification of the individual speakers in a conversation, effortlessly facilitating observation and assigning corresponding motion controls. By seamlessly integrating timestamped transcripts, customers can effortlessly jump to specific points within a recording, significantly reducing the time spent searching and ensuring that crucial details aren’t lost in translation?
- In the realm of gross sales, grasping the nuances of buyer sentiment holds paramount importance. Firms are leveraging our sentiment analysis feature to gain instant insights into customer reactions during sales calls or product demos. By incorporating time-stamped transcripts, groups are able to revisit pivotal moments in a conversation, enabling them to fine-tune their pitch and better address customer concerns in a more effective manner. In this specific instance, Named Entity Recognition (NER) plays a crucial role in identifying names, company details, and other valuable information gleaned from sales calls, enabling seamless integration with the CRM system for automated data feeding.
- Corporations operating within the heart of their contracts leverage our Application Programming Interface (API) to provide real-time assistance to brokers, while simultaneously monitoring customer sentiment during conversations. Speaker diarization enables accurate attribution of spoken comments, while time-stamped transcripts facilitate swift evaluation of critical moments or compliance points by supervisors. By elevating on-call decision-making efficiency and quality monitoring, this solution not only enhances client knowledge but also optimizes agent performance and job fulfillment.
Industries rely heavily on specialized vocabularies, model designations, and distinct linguistic subtleties. By incorporating tailored vocabulary solutions, speech-to-text technology can flexibly conform to specific requirements, thereby enabling it to accurately capture subtle contextual undertones and deliver outputs that seamlessly align with a company’s unique needs. By compiling a repository of industry-specific terminology, akin to model names, within a specific linguistic context.
Here is the rewritten text:
This adaptability enables more accurate transcripts, ultimately resulting in higher levels of professional expertise among transcribers. In many fields, such as medicine and finance, this attribute proves extremely crucial.
Named entity recognition (NER) extracts and identifies crucial information from unstructured audio data, analogous to the names of individuals, organizations, locations, and more. One common challenge with unorganized data is that vital details are hidden within the text and difficult to extract in a useful form.
To effectively address this challenge, Gladia designed and implemented a systematic Key Knowledge Extraction (KDE) approach.
By harnessing the power of Whisper’s generative architecture, similar to that of large language models (LLMs), Gladia’s Knowledge Discovery Engine (KDE) rapidly identifies and extracts relevant information by capturing contextual nuances.
This course of action will be further enhanced by offering customised vocabulary and Named Entity Recognition (NER), enabling organisations to populate Customer Relationship Management systems with key information quickly and efficiently.
Real-time transcription is revolutionizing industries in a profound manner, yielding remarkable productivity gains and tangible business benefits alongside.
Real-time transcription revolutionizes support groups by. Real-time assistance is crucial for optimizing decision-making by facilitating swift responses, informed brokerages, and superior outcomes through mechanisms such as NSF and case studies. As artificial intelligence (AI)-powered speech recognition systems become increasingly adept at processing non-English languages in real-time, customer contact centers can now deliver a truly globalized customer experience at reduced operational margins.
Pace and pinpoint accuracy in gross sales are everything. Real-time transcription empowers name brokers with timely insights, allowing them to focus on key factors that impact closing deals most significantly.
While real-time transcription may seem distant from the concerns of creatives at first glance, it nonetheless holds significant potential, particularly when it comes to live captioning and translation during media events. Although most current media applications still prioritise asynchronous transcription, this preference stems from the fact that timing is less crucial in these contexts, whereas precision is paramount for tasks such as timestamped video editing and subtitle generation.
This revolutionary technology, dubbed real-time AI, is poised to permeate every aspect of our lives. Primarily, our discussions focus on the harmonious collaboration between machines and humans, mirroring how people typically interact with one another seamlessly.
While examining the plot of a future-focused Hollywood film like Her, it’s rare to witness characters engaging with AI systems using traditional keyboard interfaces. As a testament to humanity’s collaborative potential, it is clear that voice will forever serve as the primary conduit through which we interact with our surroundings and shape the world around us.
As the primary medium for conveying and sharing human experiences, voice has played a vital role in human tradition and history for far longer than written communication. As writing emerged, it allowed us to safeguard our data more effectively by relying less on group elders to serve as custodians of our stories and wisdom.
GenAI programs, capable of comprehending spoken language, generating thoughtful replies, and logging our conversations, brought a game-changing innovation to the table. Isn’t that the pinnacle of all our endeavors? This unique form of vocal expression empowers us with a distinct sense of energy and recall, an ability previously reserved for written forms of communication to preserve memories for us. That’s why I assume our ultimate shared vision will be in disarray – it represents our collective final aspiration.
