From "output" to "input" of AI voice, what are investors betting tens of millions of dollars on?

On July 16th, voice input startup Willow Voice announced the completion of a $4.2 million angel round of financing, led by YC. Just a few days ago, on June 25th, another voice input startup, Wispr Flow, also announced the completion of a $30 million Series A round of financing.

Previously, we had been observing the AI voice track. However, most of the companies that received financing were focused on voice synthesis, that is, "output." For example, ElevenLabs, a leading company in the track, completed a $250 million Series C round of financing in January this year, with a valuation exceeding $3 billion.

However, these two recent rounds of financing seem to be sending another signal. Voice startups that focus on "input" are attracting the attention of capital.

Voice input has been around since 2012. Why can it still receive financing?

Willow Voice and Wispr Flow (hereinafter referred to as Willow and Flow for short) focus on ASR technology (Automatic Speech Recognition). The two products are quite similar. They are both somewhat like "voice input methods." Users only need to press a specific button on their computer or mobile phone to directly transcribe the spoken content into text.

At first glance, this is a function that we are already familiar with in our daily lives. For example, WeChat launched the "voice-to-text function" on the iOS platform in 2019. Apple even launched the initial version of "Voice Dictation" in iOS 6 in 2012. Moreover, star startups in the AI era such as ElevenLabs and OpenAI also cover the scenario of voice-to-text conversion.

When voice is used as input, the error rate of outputting formatted text (left), and the error rate of outputting unformatted text (right). Note: The error rate is in percentage. For example, the error rate of OpenAI's Whisper in formatted text is 14.9%. A lower data value means a stronger model ability. This test also includes scenarios such as noisy environments, strong accents, and voices with professional terms for each model. The release time of the test results: February 2025 | Image source: Voice Writer.io

Formatted: The model needs to directly output the correct format, which means not only the vocabulary is correctly recognized, but also the capitalization and punctuation are correctly output. Unformatted: Only the recognition accuracy of the words themselves is considered.

According to the test by VoiceWriter.io, except for the transcription function of Google Cloud, which is slightly worse, there is not much difference among other products. In the scenario of unformatted text, the error rate of most products is below 10%. This is comparable to the level of humans without professional transcription training. However, in the scenario of formatted text where punctuation and capitalization need to be considered, the performance of voice transcription models is slightly worse, and the average error rate of all products increases by 10%.

According to Tanay Kothari, the founder of Flow, in a podcast, although the WER of AI when transcribing unformatted text is already very low, it doesn't really matter. Because even if a product can achieve an error rate below 1%, it still means there is a wrong word in every few sentences, and users still can't fully trust AI.

Moreover, due to the differences between spoken and written language, even if the model can transcribe exactly what the user says, users won't directly send the output text as a message or save it in a note-taking app. They still need to simplify and correct the text.

Guided by this concept, the difference between Flow and traditional voice-to-text conversion is that it pursues "zero-editing information." In practice, both products add a "text processing" step between "AI directly transcribing content" and "outputting content" to provide users with text that can be directly used. This text processing step has three levels. First, formatting the text output, such as correct sentence segmentation and removing filler words. Second, understanding the context, such as automatically correcting slips of the tongue and recognizing emotions. Third, context recognition, that is, the ability to output text in different styles in different input scenarios such as DMs, emails, and notes.

Through a preliminary comparative test of Flow, Willow, and OpenAI Whisper, it was found that OpenAI Whisper's output only reaches the first level, Flow and Willow can reach the second level, and none of them can reach the third level.

The goal is great, but the reality falls short

From a product logic perspective, Flow and Willow actually complete the process from "spoken language input" to "written language output." Since it mainly focuses on written language, its usage scenarios are more inclined towards office scenarios.

In a16z's year-end AI product inventory, Ammaar Reshi, the chief designer of ElevenLabs, and entrepreneur Ben Tossell both recommended Flow. Judging from their recommendations, they use it almost every day. | Image source: a16z

Since voice input has a greater impact on the surroundings compared to keyboard input, it is not very suitable for ordinary office workers sitting at their desks. So, according to the sharing of Flow's founder, they initially targeted Silicon Valley VCs, entrepreneurs, and executives who receive a large amount of information, have a need for efficient input, and mostly have their own offices or often work outside the office.

Typical user analysis on Flow's official website | Image source: Flow's official website

After the initial user growth through VCs, entrepreneurs, and executives, Flow started to reach more users with needs through Product Hunt, such as students, code developers, creators/writers, lawyers, and consultants. Like VCs, entrepreneurs, and executives, these users also have a need to process a large amount of text or input long texts, and their work locations are generally flexible. They also often process text outside the office.

Considering the two characteristics of "text input during work" and "non-office environment," we set up three scenarios: To-do List, email reply, and pre-meeting memo, and conducted a comparative test on Willow, Flow, and ChatGPT's dictation function (powered by the Whisper model).

Test 1: To-do List scenario

Scenario description: On the way to the office by car, a team leader needs to sort out the important tasks of the day and record them in a note-taking app.

Colloquial content: Um... Today, first, I need to update the icons on the main page and then send a launch notice before 3:30 pm. Second, I need to have a review meeting with the team at 4 pm. Also, send last week's daily report to John. Third, before 5 pm, organize the user feedback summary document. Finally, before 7 pm, send next week's schedule to the product team.

Output requirements: The key information should be correct, and the to-do items should be presented in points automatically.

Outputs of different products:

Evaluation: In this scenario, none of the three products missed core information such as time and tasks. Flow and Willow both segmented the content according to the marker words like "first/second/third/finally" in the original text, looking more like a To-do List. In terms of punctuation and format, Flow did a better job.

OpenAI's Whisper performed the worst overall. Although it added punctuation, it didn't segment the content and added redundant text at the end.

Test 2: Memo scenario with professional terms

Scenario description: Before a brokerage's earnings report review meeting, as a brokerage analyst, the user needs to briefly summarize the highlights of the earnings report, form a written memo, and share it with other team members.

Colloquial content: "Um... I just read the earnings report. Although XX had year-on-year growth this quarter, the quarter-on-quarter growth was negative. And the proportion of its subscription revenue is increasing, mainly due to the contributions of XX and XX. Also, we need to take a look at the convertible bonds with Alibaba. Is there a risk of dilution? I suggest focusing on the product structure and payment momentum. The revenue growth rate is still quite conservative."

Output requirements: The key information should be correct, professional terms should be used correctly, and the tone should be formal.

Outputs of different products:

Note: The red words are the ones with errors. The text is generated by AI and is only for testing purposes, having nothing to do with reality.

Evaluation: In the memo scenario with some professional terms, all three products made mistakes in the term "dilution risk." Willow and Whisper also made quite a few other errors. Even after I manually added the term "dilution risk" in Flow, Flow still didn't output it correctly. Overall, none of the three products can handle more professional scenarios well, but Flow did a slightly better job.

In addition, none of the three products corrected "colloquial" expressions such as "the quarter-on-quarter growth was negative," nor did they do any logical sorting. When I input the content, I split the "revenue" part into the first and third sentences, and none of the three products combined the similar content together.

Test 3: Email reply scenario to a customer

Scenario description: At the airport, the user needs to reply to a customer's inquiry email and provide suggestions.

Colloquial content: Hello. I see that you want to optimize your sales process. I think your current problem is quite typical. There isn't a very systematic screening mechanism after the initial leads come in, which leads to low efficiency when the sales team follows up. In this kind of situation, several of our previous customers have also encountered it. We usually suggest standardizing the lead scoring criteria or introducing a relatively lightweight CRM system. I'll organize a previous case of ours in the next few days and send it to you later to see if it's helpful.

Output requirements: Automatically segment the content, present it in an email format, and use a formal writing style.

Outputs of different products:

Evaluation: Both Flow and Willow started the "Hello" on a new line according to the email format. Flow did a better job in terms of segmentation. None of the three products could well modify the colloquial expressions in the input. Only Flow changed "later" to "at that time." Overall, the writing style of the email is still very colloquial, and users need to manually modify it before sending.

After the experience, although the quality of Flow and Willow is acceptable, they are still quite far from their goal of "zero editing." In the recognition of professional terms and in formal writing styles, the output doesn't meet the standards. I also tested English transcription, and the results were similar, with transcription errors also occurring.

Conclusion

Although the test results show that Flow and Willow still have a certain gap from "zero editing" in scenarios such as professional terms and formal writing styles, according to media reports, Flow has a very high user stickiness and payment rate. As of now, Wispr Flow has officially announced that its monthly user growth rate has exceeded 50%, the 6 - month retention rate of active users has reached 80%, and the payment rate is as high as 19%. Its annual revenue (from July 2024 to July 2025) has reached $3.8 million.

The fact that they haven't achieved the "zero editing" goal doesn't stop users from paying. This is because products like Flow, although they can't completely eliminate the need for user intervention, still provide a differentiated experience compared to before.

Many users on Reddit and Product Hunt said that the experience of using Flow to interact with ChatGPT or for Vibe Coding is very good. | Image source: Reddit

According to comments on Reddit and Product Hunt, in less formal input scenarios, Flow performs much better than other products and is quite satisfactory. For example, some users use Flow to interact with Cursor in natural language (as shown in the picture above) for Vibe Coding. In this scenario, users only need to press a key on the Mac keyboard and can directly talk. Although Flow still has some deficiencies in the third - level conversion to written language for specific scenarios, it performs significantly better than other products led by OpenAI's Whisper in the first two levels of formatting and context understanding.

Flow's high user stickiness and payment rate also reflect that reducing the "friction" in human - machine interaction through voice input and achieving efficiency improvement may be a feasible solution. Although neither Flow nor Willow has achieved the goal of "zero editing" in all scenarios at present, with the further improvement of large - model capabilities and the accumulation of data, there will probably be great improvements in the future.

According to the founder of Flow, if "voice input" can reach a level that users can trust in the future, it won't be long before "voice input" replaces the keyboard and becomes a new paradigm of human - machine interaction (voice operating system). "Real - world efficiency improvement + the possibility of subverting the old paradigm in the future" may be the reason why VCs are willing to invest real money in voice input.

Reference articles:

1. With an 80% retention rate and a 19% payment rate, why did this AI voice keyboard secure $56 million in financing? By Crow Intelligence

2. This AI - native voice input method, Flow, raised $30 million in Series A financing. Its experience outperforms WeChat and Sogou input methods. By Uncle Huang of AI Products

3. The Best Speech Recognition API in 2025: A Head - to - Head Comparison

The data comes from third -