Speech-to-text (STT), also known as automatic speech or voice recognition, is a type of AI technology that recognizes human speech in audio or video and transcribes it into written output. In the form of an API, it can power a variety of applications, ranging from call bots to voice assistants to AI-powered virtual meeting platforms.
The commercial landscape for speech-to-text APIs today consists of the big cloud providers Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure, the famous outlier OpenAI, and specialized contenders like Gladia, Assembly AI, and others featured below.
In fact, having enjoyed a quasi-monopoly of Big Tech in the past, the market is slowly but surely being taken over by this handful of API providers who are making AI transcription more accessible to the public while largely improving its core performance and capabilities.
Our Top 5 list will focus on this market segment, being the most suitable one for enterprise-grade clients who seek production-ready, affordable APIs that can handle large volumes without compromising quality. At the end, we outline the reasons for excluding cloud providers from the final selection.
Independent STT providers design their APIs on top of existing open-source alternatives and/or build their own models to power those. Whatever the underlying architecture, a common challenge to us all is striking a perfect balance between speed, accuracy, and cost.
In fact, speed and accuracy in ASR are inversely proportional – meaning that you can’t improve one without compromising the other, at least to some extent. Cost, in turn, is a resulting factor between the two. Ultimately, it is the engineering mastery of a specific provider in addressing this tradeoff that determines the overall quality of one’s API transcription.
When deciding on a provider, it is essential to first thoroughly consider your needs, intended use case, and, of course, the budget. We recommend focusing on the following key parameters:
Here comes our list of the best five speech-to-text API providers today, based on market research, user interviews, and surveys conducted by Gladia.
Assembly AI is many user’s default choice. Founded in San Francisco, it’s a prominent player with over 6 years in the speech-to-text space, distinguished by a variety of features that cater to different user needs.
Best known for:
Room for improvement
While Assembly AI generally provides high accuracy, some users have reported inconsistent performance when it comes to language detection, code-switching, and other language support issues. Word-level timestamps in live transcription are generally reposted as inaccurate, too.
Update: As of April 2024, Assembly AI's latest model is Universal-1, boasting a series of enhancements relative to its previous core model Conformer-2.
Deepgram is another California-based pioneer of the speech-to-text APIs market that stands out for its customizable AI models and speed.
Best known for:
Room for improvement
Due to the ASR conundrum explained earlier, Deepgram’s fast inference speed and low pricing combo may come at the cost of accuracy for both batch and live transcription. Moreover, their primary focus is on English, and while it does support other languages, it is not as versatile for languages with less extensive training data.
Speechmatics is a UK-based speech recognition company with a strong global presence.
Best known for:
Room for improvement
Speechmatics' pricing structure, including different pricing for ‘basic’ vs. ‘enhanced’ quality of transcription, can be complex and may lead to unexpected costs. Some users also report experiencing difficulties in scaling up to handle a large volume of live transcription requests, which can be a limitation for enterprise-level applications.
Rev.ai, a product of Rev.com, is a well-known speech-to-text service from the US that stands out for its use of AI- and human-generated transcripts.
Best known for:
Room for improvement
Rev.ai's human-reviewed transcription service, while accurate, can introduce delays in transcription, i.e., it may take as long as 20 minutes for languages other than English. The hybrid models also make it relatively expensive compared to fully automated options. Finally, its language support is more limited than alternatives.
Gladia is a French startup founded in 2022. You may not have seen our name on many similar lists just yet – which gives us a great excuse to include ourselves in this one. Jokes aside, we may be a newcomer, but the quality of our speech-to-text API is among the best according to clients – including at TechCrunch – who had struggled with the alternatives in the past.
Best known for:
Room for improvement
The API is currently more limited in audio intelligence features compared to other alternatives.
All in all, the market is varied. There may not be one right API or open-source alternative for all — everything depends on your use case, your ideal tradeoff between speed, quality, and price, whether you require extra audio intelligence features and many other factors.
However, we can conclude that across the five key parameters proposed below – speed, accuracy, supported languages, price, and extra features – in most cases, Big Tech providers aren’t offering the best value for your money.
Take accuracy, for instance. Whereas Big Tech providers have a WER of 10%-18%, most startups and specialized providers are within the 1-10% WER range. Moreover, their speed tends to reach 25 minutes for an hour-long piece of audio. In addition to that, their pricing tends to be significantly higher than the alternatives listed above, which makes them less suitable at scale.
While the reasons for this vary, one explanation is that ASR isn’t Big Tech’s core business; providers such as Amazon, Microsoft, and Google provide ASR services as part of a broader package in their suite. If we had to pick one, however, it appears that Microsoft Azure can provide a generally fulfilling performance for some use cases, as confirmed by our users.
As to the reasons to excluding OpenAI's open-source Whisper, you can find a detailed breakdown of the model's practical limitations for enterprise users here.
Taking stock, Gladia, Assembly AI, Deepgram, Rev.ai, and Speechmatics all offer valuable speech-to-text solutions. Understanding their advantages and limitations is key to making an informed choice based on your specific needs and priorities.
At Gladia, we built an optimized version of Whisper in the form of an API, adapted to real-life use cases and distinguished by consistent accuracy, extended language support, and state-of-the-art features, including speaker diarization and word-level timestamps. If you'd like to speak with us, feel free to book a demo or sign up directly below.
Speech-To-Text
Large Language Models (LLMs) are at the forefront of the democratization of AI and they continue to get more advanced. However, LLMs can suffer from performance issues, and produce inaccurate, misleading, or biased information, leading to poor user experience and creating difficulties for product builders.
Speech-To-Text
Over the last few years, Large Language Models (LLMs) have become accessible and transformative tools, powering everything from customer support and content generation to complex, industry-specific applications in healthcare, education, and finance.
Case Studies
In recent years, fuelled by advancements in LLMs, the numbers of AI note-takers has skyrocketed. These apps are increasingly tailored to meet the unique needs of specific user groups, such as doctors, sales teams and project managers.
