From Portfolios
Battle of the AI transcribers: Meeting.ai takes on Otter.ai
Speech-to-text technology has been around for a while, but breakthroughs in generative AI are intensifying the competition to perfect it like never before.
Suddenly, everyone seems to be claiming they offer AI-powered meeting transcription services.
Otter.ai, considered a leader in the space with over US$63 million in funds raised, launched a genAI-powered chat feature last year, touting it as superior to ChatGPT. Read.ai, which recently secured US$50 million, aims to evolve its AI solutions beyond note taking.
Yet this crowded AI transcription landscape didn’t deter Hokiman Kurniawan from launching Meeting.ai in 2023, an Indonesia-based startup that hasn’t raised any external funding at all but believes it can compete globally.
“Cocktail” of LLMs
In 2016, while still a student at the University of Indonesia, Kurniawan co-founded Bahasa.ai, which specialized in creating custom chatbots for local enterprises. He identified a significant demand for natural language processing tools tailored to the Indonesian language, which is spoken by more than 270 million people.
The startup secured a couple rounds of funding from investors like East Ventures and SMDV, the VC arm of Indonesian conglomerate Sinar Mas Group. It also attracted clients such as Bank Sinarmas, MyRepublic, and Kalbe Nutritionals.
However, by 2022, Kurniawan felt that Bahasa.ai needed to pivot. “We realized by then that we had taken the business as far as we could and needed a new product with the potential to scale beyond Indonesia,” he tells Tech in Asia in a recent interview.
ChatGPT was publicly launched near the end of that year, sparking worldwide excitement over AI-powered tools. For local entrepreneurs like Kurniawan, this mainstreaming meant that they no longer needed to explain what AI is to their customers.
In July 2023, Kurniawan chose to use the recurring revenue from Bahasa.ai – which he says was in the “low millions of dollars” – to launch Meeting.ai. Bahasa.ai continues to operate and maintain services for its existing clients.
With speech recognition technology like OpenAI’s Whisper being readily available and open-sourced, it’s fairly easy for any tech-savvy individual to build a tool that can automatically transcribe human conversations.
Kurniawan notes that relying on a single infrastructure provider like OpenAI can be risky, and it may also not cost-effective as there are cheaper alternatives such as Anthropic’s Claude. He adds that by using a mix of models, Meeting.ai achieved a similar level of accuracy while reducing costs. This “cocktail” approach helps the startup balance quality with affordability.
Moreover, based on a recent report by the Associated Press, Whisper has a high tendency to fabricate texts that are not in the conversations being transcribed.
For Meeting.ai, developing its own AI model isn’t financially feasible. And even if it could, the risk remains that competitors might quickly develop better models within weeks.
“Even OpenAI, with its vast resources, managed to hold a lead for only about a year before companies like Anthropic and Google caught up with comparable AI models,” Kurniawan points out.
The CEO believes that Meeting.ai can have more solid footing in the landscape by curating existing AI models so that the solution can adapt to various use cases. For example, after completing the transcription, Meeting.ai uses different models to check for typos, identify topics, and align them with the transcript when generating summaries.
However, Kurniawan found that existing models lacked accuracy when transcribing conversations in the Indonesian language. As a result, Meeting.ai developed its own model specializing in the language, training it with data that the startup had collected since 2017. The tool uses this model in conjunction with other options when transcribing Indonesian conversations.
The company also fine-tuned several open-source models by supplying additional training data. For instance, to better capture Singlish, the startup used over two terabytes of voice data sourced from the National Library of Singapore.
Meeting.ai is also banking on its name as a competitive advantage. Kurniawan points out that its domain name has a large potential for search engine optimization, letting many users discover the startup by searching for terms like “meeting AI” or “minutes AI.”
The firm says it has secured trademark rights for its name in the UK, with a US trademark currently pending for approval.
The AI transcriber showdown
But just how good is Meeting.ai at transcribing conversations? And more importantly, how does it stack up against more established options?
Internal tests done by Tech in Asia show that for English conversations between non-native speakers, Meeting.ai’s transcription accuracy is comparable to Otter.ai and TurboScribe.
For conversations between Indonesian speakers, Meeting.ai can recognize several common slang words such as “gue,” “lu,” “kayak,” “pede,” and “gede,” which respectively mean “I,” “you,” “like,” “confident,” and “large.” This puts it on par with Prosa.ai’s solution, another startup that specializes in natural language processing for the Indonesian language.
Common English slang words from different countries (Australia, Singapore, the UK, and the US) are also supported, according to the company.
To further fine-tune the tool to identify slang in other languages, Meeting.ai needs support from local partners, which it currently lacks.
The tool can also detect multiple languages in a single meeting – English and Indonesian, for instance – a feature that Otter.ai and Fireflies.ai are currently lacking.
Meeting.ai’s web version, however, currently lacks a real-time transcription feature, which can be a deal breaker for anyone looking for instant feedback during meetings. Additionally, the recording feature doesn’t support partial uploading, so users must wait for the entire upload process to finish after each conversation.
At this rate, it’s hard for an AI-based solution to remain unique for more than three months.
In addition to enhancing these areas, Meeting.ai is exploring the integration of features that allow users to work on tasks based on their conversations – such as drafting a term sheet or proposal – all within its app.
Kurniawan hopes these additions will help his startup remain competitive, but he is fully aware that such advantage will only be temporary. “At this rate, it’s hard for an AI-based solution to remain unique for more than three months. You have to always accept the fact that others will catch up quickly,” he says.
No urgent need for fundraising
When asked about the decision to remain bootstrapped, Kurniawan explains that the startup doesn’t urgently need external funding. However, he is open to raising around US$3 million to US$5 million in early 2025 to enhance global marketing efforts.
Currently, Meeting.ai’s team consists of 20 people, primarily engineers. It has over 111,000 total active users across the globe, with about 20,000 regularly accessing the app every month.
While the CEO didn’t disclose current revenue, he notes that 70% comes from Indonesia. The rest comes from markets like Singapore, Malaysia, the UK, and the US.
In addition to individual users, Meeting.ai also serves enterprises and government agencies such as Telkomsel, Indosat, and the National Research and Innovation Agency.
Unlike many competitors, Meeting.ai charges customers on a per-hour basis instead of per user. Individual users can opt for a $27 per month plan, which provides 10 hours of transcription credit—enough to transcribe up to 10 hours of recordings. A $49 per month plan offers double the credit, along with additional features.
“This model guarantees positive unit economics for us because it can’t be exploited through account sharing methods,” explains Kurniawan.
He adds that as of September, Meeting.ai is already EBITDA positive, and it is looking to hit net profitability “soon.”
Artikel asli telah diterbitkan di Tech in Asia, 8 November 2024.
Artikel terbaru
Rekomendasi bacaan
