15.ai

Features

15.ai was non-commercial, had no advertisements, generated no revenue, and operated without requiring user registration or accounts. Users were able to generate speech by inputting text and selecting a character voice, with optional parameters for emotional contextualizers and phonetic transcriptions. Each request produced three audio variations with distinct emotional deliveries. Characters available included multiple characters from Team Fortress 2 and My Little Pony: Friendship Is Magic, including the Mane Six and Derpy Hooves; GLaDOS, Wheatley, and the Sentry Turret from the Portal series; SpongeBob SquarePants; Kyu Sugardust from HuniePop, Rise Kujikawa from Persona 4; Daria Morgendorffer and Jane Lane from Daria; Carl Brutananadilewski from Aqua Teen Hunger Force; Steven Universe from Steven Universe; Sans from Undertale; Madeline and multiple characters from Celeste; the Tenth Doctor from Doctor Who; the Narrator from The Stanley Parable; and HAL 9000 from 2001: A Space Odyssey. Silent characters like Chell and Gordon Freeman were able to be selected and would emit silent audio files when any text was submitted. Characters from Undertale and Celeste did not produce spoken words but instead generated their games' distinctive beeps when text was entered.

From 2020, 15.ai had generated audio at 44.1 kHz sampling rate—higher than the 16 kHz standard used by most deep learning text-to-speech systems of that period. This higher fidelity created more detailed audio spectrograms and greater audio resolution with the tradeoff that imperfections in the synthesis were more noticeable. 15.ai processed speech using customized deep neural networks and specialized audio synthesis algorithms. While its underlying technology could produce 10 seconds of audio in less than 10 seconds of processing time (i.e. faster-than-real-time), the user experience often involved longer waits as the servers managed thousands of simultaneous requests, sometimes taking more than a minute to deliver results.

Due to its nondeterministic design, 15.ai produced variations in its speech output. 15.ai introduced the concept of emotional contextualizers, which allowed users to specify the emotional tone of generated speech through guiding phrases. The emotional contextualizer functionality utilized DeepMoji to inject emotional context into synthesized speech. DeepMoji was a sentiment analysis neural network developed at the MIT Media Lab that processed emoji embeddings from 1.2 billion Twitter posts. If a prompt contained additional context (specified by a vertical bar), the additional context following the bar would be used as the emotional contextualizer. For example, if the input was Today is a great day!|I'm very sad., the selected character would speak the sentence "Today is a great day!" in the emotion one would expect from someone saying the sentence "I'm very sad."

15.ai used pronunciation data from Oxford Dictionaries API, Wiktionary, and CMU Pronouncing Dictionary, which uses ARPABET phonetic transcriptions. Users could input ARPABET transcriptions by enclosing phoneme strings in curly braces to correct mispronunciations. 15.ai's interface used color-coding to indicate pronunciation certainty and also displayed technical metrics, graphs, and comprehensive model analytics, which had included sentiment analysis and automatic improvements to the vocoder. The platform limited its prompt to 200 characters; users could combine multiple generations for longer speech sequences.

Later versions of 15.ai introduced multi-speaker capabilities. Rather than training separate models for each voice, 15.ai used a unified model that learned multiple voices simultaneously through speaker embeddings: numerical representations that capture each character's unique vocal characteristics. Along with the emotional context conferred by DeepMoji, this allowed the deep learning model to learn shared patterns across different characters' emotional expressions and speaking styles, even when characters lacked examples of certain emotions in their training data.

Fan works and applications

Fan works that used 15.ai for voice acting included skits and fan animations, crossover content, recreations of viral videos, adaptations of fan fiction, music videos, and musical compositions. Team Fortress 2 content creators regularly used the platform to produce both short-form memes and complex narrative animations using Source Filmmaker.

PC Gamer showcased Team Fortress 2 characters redubbing a scene from Thomas the Tank Engine and arguing over a game of Among Us, while United Daily News showcased a remake of the Team Fortress 2 video "The Heavy Is Dead" and parodies of the viral "Lamar roasts Franklin" scene from Grand Theft Auto V. Kotaku showcased a version of the car dealership parody "Big Bill Hell's Cars" narrated by SpongeBob SquarePants, and Game Informer featured a Portal–Mass Effect crossover created by the author of the article herself. AUTOMATON showcased multiple Team Fortress 2 videos made using 15.ai. One used Source Filmmaker to animate a 15.ai-generated skit in which Demoman confronts Spy after discovering a drawing of his wolf fursona. Another redubbed the card reading scene from the Team Fortress short film Expiration Date, recasting it as Spy reviewing what users had created with 15.ai and growing exasperated as every example concerned a fursona, including his own. Some fan creations gained mainstream attention: a viral video that replaced Donald Trump's cameo in Home Alone 2: Lost in New York with the Heavy Weapons Guy's AI-generated voice was featured on a daytime CNN segment in January 2021.

15.ai was also regularly used within the My Little Pony fandom. In March 2020, an Equestria Daily event for generated pony voices received dozens of submissions, including episode redubs, advertisement parodies, literary recitations, song and rap covers, crossovers, and original skits. Later works featured by Equestria Daily included a full playthrough of the horror fan game Story of the Blanks using 15.ai voices, the viral Friendship is Magic–Among Us crossover animation "Among Us Struggles", and "The Tax Breaks", a fully voiced 17-minute fan-made episode.

Some users created virtual assistants using voices generated with 15.ai. In one documented project, a Portal fan used 15.ai to provide GLaDOS's voice for a customized Amazon Echo housed in a 3D printed realistic replica of the character's head, which consisted of more than 200 specially designed parts. The virtual assistant was programmed to answer selected requests with 15.ai-generated GLaDOS dialogue. Separately, Denfaminico Gamer reported a fan-created demonstration that combined 15.ai with the voice command software VoiceAttack to use GLaDOS as an assistant for controlling a computer.

Reception

Critical reception

Critics described 15.ai as easy to use and generally able to convincingly replicate character voices, with occasional mixed results. Liana Ruppert of Game Informer described 15.ai as "simplistically brilliant" and praised its ease of use. Natalie Clayton of PC Gamer wrote that SpongeBob SquarePants' voice was replicated well, but described challenges in mimicking the Narrator from The Stanley Parable: "the algorithm simply can't capture Kevan Brighting's whimsically droll intonation." Zack Zwiezen of Kotaku reported that "[his] girlfriend was convinced it was a new voice line from GLaDOS' voice actor". Taiwanese newspaper United Daily News also highlighted 15.ai's ability to recreate GLaDOS's mechanical voice, alongside its diverse range of character voice options. Marco Cocomello of GLITCHED remarked that despite the 200-character limitation, the results "blew [him] away" when testing the app with GLaDOS's voice.

Carl Johansson-Sundelius of FZ wrote that the underlying technology of 15.ai was "impressive, to say the least", and acknowledged the technical breadth of the various models used to produce the voices. Chris Button of Byteside called the ability to clone a voice with only 15 seconds of data "freaky," and also described the tech behind it as "impressive." Robin Lamorlette of Clubic described the technology as "devilishly fun" and wrote that Twitter and YouTube were filled with creative content from users experimenting with the tool. The platform's voice generation capabilities were regularly featured on Equestria Daily with documented updates, fan creations, and additions of new character voices. In a post introducing new character additions to 15.ai, Equestria Daily's founder Shaun Scotellaro wrote that "some of [the voices] aren't great due to the lack of samples to draw from, but many are really impressive still anyway." Chinese My Little Pony fan site EquestriaCN also documented 15.ai's development and its updates, though they criticized some of its bugs and long queue wait times.

Peter Paltridge of Anime Superhero News opined that "voice synthesis has evolved to the point where the more expensive efforts are nearly indistinguishable from actual human speech," but also stated that "In some ways, SAM is still more advanced than this. It was possible to affect SAM's inflections by using special characters, as well as change his pitch at will. With 15.ai, you're at the mercy of whatever random inflections you get." Conversely, Lauren Morton of Rock, Paper, Shotgun praised the depth of pronunciation control—"if you're willing to get into the nitty gritty of it". Similarly, Eugenio Moto of Qore.com wrote that "the most experienced of users can change parameters like the stress or the tone." Spanish author Álvaro Ibáñez wrote in Microsiervos that he found the rhythm of the AI-generated voices "interesting" and that 15.ai was able to adapt its delivery based on the text's meaning. Amaya Tomanová of Letem světem Applem described 15.ai's voice generation as "lightning fast" and highlighted the ability to use emojis as part of the prompt, also calling it "interesting".

Multiple reporters criticized the lack of non-English language support. Takayuki Furushima of Den Fami Nico Gamer highlighted the "smooth pronunciations", and Yuki Kurosawa of AUTOMATON wrote that its "rich emotional expression" was a major feature; both Japanese authors mentioned the lack of Japanese-language support. Renan do Prado of Arkade and José Villalobos of LaPS4 remarked that while users could create amusing results in Portuguese and Spanish respectively, the generation performed best in English. Chinese gaming news website GamerSky called the app "interesting", but also criticized the word count limit of the text and the occasional lack of intonations. Yahoo! News Taiwan reported that "GLaDOS in Portal can pronounce lines nearly perfectly", but also criticized that "there are still many imperfections, such as word limit and tone control, which are still a little weird in some words."

Technical reception

AI researchers and practitioners provided more in-depth analysis of 15.ai's capabilities and limitations compared to other text-to-speech technologies of the time. Google DeepMind senior research scientist Alex Irpan wrote that when 15.ai launched in 2020, it was "arguably the highest quality voice generation model in the world" and superior to models developed by Google AI. In their explanation for why they plagiarized from 15.ai, Voiceverse wrote that "the best voice [their marketing team] could find was not [theirs] but 15.ai['s]". Rionaldi Chandraseta of Towards Data Science wrote that voice models trained on larger datasets created more convincing output with better phrasing and natural pauses, particularly for extended text. Parth Mahendra of AI Daily wrote that while the system "does a good job at accurately replicating most basic words," it struggled with more complex terms, noting that characters would "absolutely butcher the pronunciation" of certain words. Dan Ruta, the creator of a downloadable voice AI tool called xVASynth, wrote that in comparison, cloud-based tools like 15.ai had the benefit of being accessible on any device. Economist and AI commentator Tyler Cowen named 15.ai among examples of underrated talent in AI and machine learning.

Chinese AI researchers also offered various technical assessments. Machine learning professor Yongqiang Li wrote that 15.ai "perfectly preserves the rhythm and characteristics of the speaker," and remarked that the application was still free despite having 5,000 people generating voices concurrently at the time of writing. Bai Feng of XinZhiYuan on QQ News highlighted the technical achievement of 15.ai's high-quality output despite using minimal training data and wrote that it was of significantly higher quality than typical deep learning text-to-speech implementations. Feng also acknowledged that while some pronunciation errors occurred due to the limited training data, it was understandable given that contemporary deep learning models typically required 40 or more hours of audio. Ji Yunyo of NetEase News called the technology behind 15.ai "remarkably efficient" but also criticized its emotional limitations, writing that the emotional expression was relatively "neutral" and that "extreme" emotions couldn't be properly synthesized, making it less suitable for not safe for work applications. Ji also wrote that while many deepfake videos required creators to extract and edit material from hours of original content for very short results, 15.ai could achieve similar or better effects with only a few dozen minutes of training data per character.

Reactions from voice actors of featured characters

Some voice actors whose characters appeared on 15.ai have publicly shared their thoughts about the platform. In an April 2021 interview, John Patrick Lowrie—who voices the Sniper in Team Fortress 2—said that he had discovered 15.ai when a prospective intern showed him a skit she had created using AI voices of the Team Fortress 2 characters. Lowrie commented:

"The technology still has a long way to go before you really believe that these are just human beings, but I was impressed by how much [15.ai] could do. [...] You certainly don't get the delivery that you get from an actual person who's analyzed the scene, [...] but I do think that as a fan source—for people wanting to put together mods and stuff like that—that it could be fun for fans to use the voices of characters they like."

He drew an analogy to synthesized music, adding:

"If you want the sound of a choir, and you want the sound of an orchestra, and you have the money, you hire a choir and an orchestra. And if you don't have the money, you have something that sounds pretty nice; but it's not the same as a choir and an orchestra."

In a 2021 live broadcast on his Twitch channel, Nathan Vetterlein—the voice actor of the Scout from Team Fortress 2—listened to an AI recreation of his character's voice and commented: "It's interesting; it's all right. There's some stuff in there".

Ethical concerns

Other voice actors had mixed reactions to 15.ai's capabilities. While some industry professionals acknowledged the technical innovation, others raised concerns about the technology's implications for their profession. When voice actor Troy Baker announced his partnership with Voiceverse NFT, which had misappropriated 15.ai's technology, critics raised concerns about automated voice acting's potential reduction of employment opportunities for voice actors, risk of voice impersonation, and potential misuse in explicit content. Ruby Innes of Kotaku Australia wrote that "this practice could potentially put voice actors out of work considering you could just use their AI voice rather than getting them to voice act for a project and paying them." In her coverage of the Voiceverse controversy, Edith Wakelin-King of Checkpoint Gaming raised the concern that "this kind of technology has the potential to push voice actors out of work if it becomes easier and cheaper to use AI voices instead of working with the actor directly."

While 15.ai limited its scope to fictional characters and did not reproduce voices of real people or celebrities, computer scientist Andrew Ng commented that similar technology could be used to do so, including for nefarious purposes. In his 2020 assessment of 15.ai, Ng outlined potential "enormously productive" applications of voice cloning, such as revolutionizing the use of virtual actors in Hollywood, enabling voice actors to participate in more cartoon and audiobook productions, and allowing content creators to use synthetic celebrity voices to narrate their scripts. However, he also cautioned that synthesizing a human's voice without consent raises ethical concerns and potential legal issues, and further warned that it could be maliciously exploited to impersonate private individuals.

Legacy

15.ai was one of the first mainstream applications of generative artificial intelligence and was the first mainstream application of audio deepfake technology. 15.ai is credited with popularizing AI voice cloning in Internet memes and content creation, particularly through its ability to generate convincing character voices in real time without requiring extensive technical expertise. The platform's impact was especially large in fan communities, such as the My Little Pony: Friendship Is Magic, Portal, Team Fortress 2, and SpongeBob SquarePants fandoms, where it enabled the creation of viral content that garnered millions of views on social media.

The popularity and success of 15.ai made its creator, 15, a prominent figure. Google DeepMind senior research scientist Alex Irpan described 15 as one of the most famous people in the brony fandom (a status known as being "horse famous" in the vernacular of the brony fandom), attributing this status to 15.ai's high quality, its free availability, and 15's willingness to personally cover thousands of dollars in monthly operating costs. Its influence since its launch has been publicly recognized, with commercial alternatives like ElevenLabs and Speechify emerging to fill the void after its initial shutdown. Multiple contemporary generative voice AI companies have acknowledged 15.ai's pioneering role.

15.ai established technical precedents that influenced subsequent developments in AI voice synthesis and was one of the most serious and consequential results of early experimentation with incorporating AI into voice synthesis. Its integration of DeepMoji for emotional analysis demonstrated the viability of incorporating sentiment-aware speech generation, while its support for ARPABET phonetic transcriptions set a standard for precise pronunciation control in public-facing voice synthesis tools. The platform's multi-speaker model, which enabled simultaneous training of diverse character voices, allowed the system to recognize emotional patterns across different voices even when certain emotions were absent from individual character training sets. 15.ai also contributed to the reduction of training data requirements for speech synthesis. Contemporary models like Tacotron 2 required tens of hours of audio to produce acceptable results and failed to generate intelligible speech with less than 24 minutes of training data. In contrast, 15.ai demonstrated the ability to generate speech with substantially less training data; the name "15.ai" refers to the creator's statement that a voice can be cloned with just 15 seconds of data. The 15-second benchmark became a reference point for subsequent voice synthesis systems; the original statement that only 15 seconds of data is required to clone a human's voice was corroborated by OpenAI in 2024.