Jean-Louis Quéguiner is the Founder and CEO of Gladia. He beforehand served as Group Vice President of Information, AI, and Quantum Computing at OVHcloud, one in all Europe’s main cloud suppliers. He holds a Grasp’s Diploma in Symbolic AI from the College of Québec in Canada and Arts et Métiers ParisTech in Paris. Over the course of his profession, he has held vital positions throughout numerous industries, together with monetary information analytics, machine studying purposes for real-time digital promoting, and the event of speech AI APIs.
Gladia supplies superior audio transcription and real-time AI options for seamless integration into merchandise throughout industries, languages, and expertise stacks. By optimizing state-of-the-art ASR and generative AI fashions, it ensures correct, lag-free speech and language processing. Gladia’s platform additionally allows real-time extraction of insights and metadata from calls and conferences, supporting key enterprise use circumstances corresponding to gross sales help and automatic buyer help.
What impressed you to sort out the challenges in speech-to-text (STT) expertise, and what gaps did you see available in the market?
After I based Gladia, the preliminary purpose was broad—an AI firm that might make complicated expertise accessible. However as we delved deeper, it grew to become clear that voice expertise was probably the most damaged and but most important space to concentrate on.
Voice is central to our each day lives, and most of our communication occurs by speech. But, the instruments accessible for builders to work with voice information have been insufficient when it comes to pace, accuracy, and value—particularly throughout languages.
I needed to repair that, to unpack the complexity of voice expertise and repackage it into one thing easy, environment friendly, highly effective and accessible. Builders shouldn’t have to fret in regards to the intricacies of AI fashions or the nuances of context size in speech recognition. My purpose was to create an enterprise-grade speech-to-text API that labored seamlessly, whatever the underlying mannequin or expertise—a real plug-and-play answer.
What are a number of the distinctive challenges you encountered whereas constructing a transcription answer for enterprise use?
In the case of speech recognition, pace and accuracy—the 2 key efficiency indicators on this area—are inversely proportional by design. Because of this enhancing one will compromise the opposite, at the very least to some extent. The associated fee issue, to a giant extent, outcomes from the supplier’s selection between pace and high quality.
When constructing Gladia, our purpose was to seek out the proper steadiness between these two elements, all whereas making certain the expertise stays accessible to startups and SMEs. Within the course of we additionally realized that the foundational ASR fashions like OpenAI’s Whisper, which we labored with extensively, are biased, skewering closely in direction of English as a consequence of their coaching information, which leaves a variety of languages under-represented.
So, along with fixing the speed-accuracy tradeoff, it was necessary to us— as a European, multilingual crew—to optimize and fine-tune our core fashions to construct a really international API that helps companies function throughout languages.
How does Gladia differentiate itself within the crowded AI transcription market? What makes your Whisper-Zero ASR distinctive?
Our new real-time engine (Gladia Actual Time) achieves an industry-leading 300 ms latency. Along with that, it’s in a position to extract insights from a name or assembly with the so-called “audio intelligence” add-ons or options, like named entity recognition (NER) or sentiment evaluation.
To our information, only a few opponents are in a position to present each transcription and insights at such excessive latency (lower than 1s end-to-end) – and do all of that precisely in languages apart from English. Our languages help extends to over 100 languages immediately.
We additionally put a particular emphasis on making the product really stack agnostic. Our API is appropriate with all present tech stacks and telephony protocols, together with SIP, VoIP, FreeSwitch and Asterisk. Telephony protocols are particularly complicated to combine with, so we imagine this product side can convey great worth to the market.
Hallucinations in AI fashions are a big concern, particularly in real-time transcription. Are you able to clarify what hallucinations are within the context of STT and the way Gladia addresses this drawback?
Hallucination often happens when the mannequin lacks information or doesn’t have sufficient context on the subject. Though fashions can produce outputs tailor-made to a request, they’ll solely reference data that existed on the time of their coaching, and that might not be up-to-date. The mannequin will create coherent responses by filling in gaps with data that sounds believable however is wrong.
Whereas hallucinations grew to become recognized within the context of LLMs first, they happen with speech recognition fashions— like Whisper ASR, a number one mannequin within the area developed by OpenAI – as properly. Whisper’s hallucinations are like these of LLMs as a consequence of an identical structure, so it’s an issue that issues generative fashions, which can be in a position to predict the phrases that observe based mostly on the general context. In a means, they ‘invent’ the output. This strategy might be contrasted with extra conventional, acoustic-based ASR architectures that match the enter sound to output in a extra mechanical means
In consequence, chances are you’ll discover phrases in a transcript that weren’t really mentioned, which is clearly problematic, particularly in fields like drugs, the place a mistake of this type can have grave penalties.
There are a number of strategies to handle and detect hallucinations. One frequent strategy is to make use of a retrieval-augmented era (RAG) system, which mixes the mannequin’s generative capabilities with a retrieval mechanism to cross-check details. One other methodology includes using a “chain of thought” strategy, the place the mannequin is guided by a collection of predefined steps or checkpoints to make sure that it stays on a logical path.
One other technique for detecting hallucinations includes utilizing methods that assess the truthfulness of the mannequin’s output throughout coaching. There are benchmarks particularly designed to guage hallucinations, which contain evaluating totally different candidate responses generated by the mannequin and figuring out which one is most correct.
We at Gladia have experimented with a mixture of methods when constructing Whisper-Zero, our proprietary ASR that removes just about all hallucinations. It’s confirmed wonderful ends in asynchronous transcription, and we’re presently optimizing it for real-time to attain the identical 99.9% data constancy.
STT expertise should deal with a variety of complexities like accents, noise, and multi-language conversations. How does Gladia strategy these challenges to make sure excessive accuracy?
Language detection in ASR is an especially complicated job. Every speaker has a novel vocal signature, which we name options. By analyzing the vocal spectrum, machine studying algorithms can carry out classifications, utilizing the Mel Frequency Cepstral Coefficients (MFCC) to extract the primary frequency traits.
MFCC is a technique impressed by human auditory notion. It’s a part of the “psychoacoustic” area, specializing in how we understand sound. It emphasizes decrease frequencies and makes use of methods like normalized Fourier decomposition to transform audio right into a frequency spectrum.
Nevertheless, this strategy has a limitation: it is based mostly purely on acoustics. So, if you happen to converse English with a powerful accent, the system might not perceive the content material however as a substitute choose based mostly in your prosody (rhythm, stress, intonation).
That is the place Gladia’s modern answer is available in. We have developed a hybrid strategy that mixes psycho-acoustic options with content material understanding for dynamic language detection.
Our system would not simply take heed to the way you converse, but in addition understands what you are saying. This twin strategy permits for environment friendly code-switching and would not let robust accents get misrepresented/misunderstood.
Code-switching—which is amongst our key differentiators—is a very necessary function in dealing with multilingual conversations. Audio system might change between languages mid-conversation (and even mid-sentence), and the flexibility of the mannequin to transcribe precisely on the fly regardless of the change is essential.
Gladia API is exclusive in its capacity to deal with code-switching with this many language pairs with a excessive stage of accuracy and performs properly even in noisy environments, recognized to cut back the standard of transcription.
Actual-time transcription requires ultra-low latency. How does your API obtain lower than 300 milliseconds latency whereas sustaining accuracy?
Retaining latency underneath 300 milliseconds whereas sustaining excessive accuracy requires a multifaceted strategy that blends {hardware} experience, algorithm optimization, and architectural design.
Actual-time AI isn’t like conventional computing—it’s tightly linked to the ability and effectivity of GPGPUs. I’ve been working on this house for practically a decade, main the AI division at OVHCloud (the largest cloud supplier within the EU), and discovered firsthand that it’s at all times about discovering the precise steadiness: how a lot {hardware} energy you want, how a lot it prices, and the way you tailor the algorithms to work seamlessly with that {hardware}.
Efficiency in actual time AI comes from successfully aligning our algorithms with the capabilities of the {hardware}, making certain each operation maximizes throughput whereas minimizing delays.
However it’s not simply the AI and {hardware}. The system’s structure performs a giant function too, particularly the community, which might actually influence latency. Our CTO, who has deep experience in low-latency community design from his time at Sigfox (an IoT pioneer), has optimized our community setup to shave off useful milliseconds.
So, it’s actually a mixture of all these elements—sensible {hardware} selections, optimized algorithms, and community design—that lets us constantly obtain sub-300ms latency with out compromising on accuracy.
Gladia goes past transcription with options like speaker diarization, sentiment evaluation, and time-stamped transcripts. What are some modern purposes you’ve seen your purchasers develop utilizing these instruments?
ASR unlocks a variety of purposes to platforms throughout verticals, and it’s been superb to see what number of really pioneering corporations have emerged within the final two years, leveraging LLMs and our API to construct cutting-edge, aggressive merchandise. Listed here are some examples:
- Good note-taking: Many purchasers are constructing instruments for professionals who have to rapidly seize and set up data from work conferences, scholar lectures, or medical consultations. With speaker diarization, our API can establish who mentioned what, making it simple to observe conversations and assign motion objects. Mixed with time-stamped transcripts, customers can leap straight to particular moments in a recording, saving time and making certain nothing will get misplaced in translation.
- Gross sales enablement: Within the gross sales world, understanding buyer sentiment is every thing. Groups are utilizing our sentiment evaluation function to achieve real-time insights into how prospects reply throughout calls or demos. Plus, time-stamped transcripts assist groups revisit key elements of a dialog to refine their pitch or handle shopper issues extra successfully. For this use case specifically, NER can be key to figuring out names, firm particulars, and different data that may be extracted from gross sales calls to feed the CRM mechanically.
- Name middle help: Firms within the contract middle house are utilizing our API to supply dwell help to brokers, in addition to flagging buyer sentiment throughout calls. Speaker diarization ensures that issues being mentioned are assigned to the precise particular person, whereas time-stamped transcripts allow supervisors to evaluate essential moments or compliance points rapidly. This not solely improves the client expertise – with higher on-call decision charge and high quality monitoring – but in addition boosts agent productiveness and satisfaction.
Are you able to talk about the function of customized vocabularies and entity recognition in enhancing transcription reliability for enterprise customers?
Many industries depend on specialised terminology, model names, and distinctive language nuances. Customized vocabulary integration permits the STT answer to adapt to those particular wants, which is essential for capturing contextual nuances and delivering output that precisely displays your online business wants. As an illustration, it permits you to create a listing of domain-specific phrases, corresponding to model names, in a particular language.
Why it’s helpful: Adapting the transcription to the precise vertical permits you to decrease errors in transcripts, attaining a greater person expertise. This function is very essential in fields like drugs or finance.
Named entity recognition (NER) extracts and identifies key data from unstructured audio information, corresponding to names of individuals, organizations, areas, and extra. A standard problem with unstructured information is that this essential data isn’t readily accessible—it is buried inside the transcript.
To resolve this, Gladia developed a structured Key Information Extraction (KDE) strategy. By leveraging the generative capabilities of its Whisper-based structure—much like LLMs—Gladia’s KDE captures context to establish and extract related data immediately.
This course of might be additional enhanced with options like customized vocabulary and NER, permitting companies to populate CRMs with key information rapidly and effectively.
In your opinion, how is real-time transcription remodeling industries corresponding to buyer help, gross sales, and content material creation?
Actual-time transcription is reshaping these industries in profound methods, driving unimaginable productiveness positive factors, coupled with tangible enterprise advantages.
First, real-time transcription is a game-changer for help groups. Actual-time help is essential to enhancing the decision charge because of sooner responses, smarter brokers, and higher outcomes (when it comes to NSF, deal with instances, and so forth). As ASR methods get higher and higher at dealing with non-English languages and performing real-time translation, contact facilities can obtain a really international CX at decrease margins.
In gross sales, pace and spot-on insights are every thing. Equally to what occurs with name brokers, real-time transcription is what equips them with the precise insights on the proper time, enabling them to concentrate on what issues probably the most in closing offers.
For creators, real-time transcription is maybe much less related immediately, however nonetheless filled with potential, particularly in the case of dwell captioning and translation throughout media occasions. Most of our present media prospects nonetheless choose asynchronous transcription, as pace is much less essential there, whereas accuracy is essential for purposes like time-stamped video enhancing and subtitle era.
Actual-time AI transcription appears to be a rising pattern. The place do you see this expertise heading within the subsequent 5-10 years?
I really feel like this phenomenon, which we now name real-time AI, goes to be in all places. Basically, what we actually discuss with right here is the seamless capacity of machines to work together with individuals, the way in which we people already work together with each other.
And if you happen to take a look at any Hollywood film (like Her) set sooner or later, you’ll by no means see anybody there interacting with clever methods by way of a keyboard. For me, that serves as the final word proof that within the collective creativeness of humanity, voice will at all times be the first means we work together with the world round us.
Voice, as the primary vector to combination and share human information, has been a part of human tradition and historical past for for much longer than writing. Then, writing took over as a result of it enabled us to protect our information extra successfully than counting on the neighborhood elders to be the guardians of our tales and knowledge.
GenAI methods, able to understanding speech, producing responses, and storing our interactions, introduced one thing fully new to the house. It’s one of the best of each phrases and one of the best of humanity actually. It offers us this distinctive energy and power of voice communication with the advantage of reminiscence, which beforehand solely written media might safe for us. Because of this I imagine it’s going to be in all places – it is our final collective dream.
Thanks for the nice interview, readers who want to study extra ought to go to Gladia.