Mikey Shulman

Mikey Shulman is the co-founder and CEO of Suno, where he has advocated transformer-based audio generation and tools that let ordinary listeners become music makers. His technical priorities have included learned audio representations, limited imposed musical knowledge and models fast enough to stream music at listening speed. On September 9, 2026, he announced v6, a generation of music models developed with industry partners including Warner Music Group, BMG and Believe.[1][6]

Before Suno, Shulman worked on machine learning for finance at Kensho and taught at MIT Sloan. That earlier work provides context for the text and speech-recognition projects he shared with his future co-founders.[3][1]

Shulman earned a doctorate in physics from Harvard. By November 2020, he was Kensho’s head of machine learning and a finance lecturer three years into teaching at MIT Sloan. He taught Advanced Data Analytics and Machine Learning in Finance with Georg Kucsko, then Kensho’s head of machine learning research and development.[3]

In his 2020 explanation of natural language processing, Shulman described how financial information arrives as unstructured text and speech. His example of searching 10-K filings for Obamacare illustrated the limits of literal matching: a useful system would need to recognize the formal term PPACA. He also argued that accessible, open-source tools made advanced NLP feasible without a doctorate in machine learning.[3]

Shulman and Suno co-founders Georg, Martin and Keenan previously worked together at Kensho. Recalling the transition in March 2024, he said their work was mostly text until a financially focused speech-recognition project drew them into audio. All four were musicians and music lovers. Although advisers urged them toward the larger speech market, they kept experimenting with music. He described the choice as organic rather than a single breakthrough: “I don't think there was necessarily like an aha moment.”[1]

The team’s first release was Bark, an open-source transformer-based text-to-audio model whose repository records an April 20, 2023 release and a May 1 move to the MIT License. Shulman described training a transformer through self-supervised learning on broad audio, then adapting it for text-to-speech. He credited Andrej Karpathy’s nanoGPT code and identified the team’s contribution as turning audio into tokens. Bark’s documentation also credits EnCodec, AudioLM, Vall-E and preceding research; Shulman separately identified Parakeet as a collaboration with NVIDIA.[1][2]

In his March 2024 interview, Shulman explained the transformer approach as predicting successive audio tokens, with learned encoders converting audio into discrete representations. He favored imposing as little musical knowledge as possible, accepting potential short-term disadvantages so the model could learn more general behavior. Bark demonstrates one concrete version of that choice: it converts text directly to audio without intermediate phonemes, allowing music and non-speech sounds as well as speech. Its generative output can also depart from the supplied script.[1][2]

Shulman linked model size to a practical constraint: music must be generated at least as quickly as listeners consume it. He said the team therefore preferred relatively small models where possible, and that early compute constraints forced research they might have skipped with larger models. He described three evaluation axes—sound fidelity, song quality and controllability—including testing the same lyrics in different styles. His broader principle was “aesthetics matter”: quantitative benchmarks could help, but listening remained necessary to judge music’s emotional effect.[1]

In March 2024, Shulman said Suno’s priority was changing how ordinary people participate in music, rather than primarily improving professional production workflows. He acknowledged professionals already using it for inspiration and samples. He reported that more than half of usage involved expert mode, where users could tweak lyrics and other details, alongside quick creations prompted by everyday events.[1]

One example was his recollection of a coffee order: co-founder Martin’s name came back as Margoo, and Shulman turned the mistake into a song that stayed in the team’s heads. He saw such small, shared moments as a meaningful use of music generation. He also imagined collaborative concerts in which audience members could articulate the sounds in their heads through AI, while making clear that this was an idea for the future.[1]

Shulman emphasized original music in 2024 and said Suno did not allow remaking songs with someone else’s lyrics. In his August 6, 2026 statement, he said the company excluded artist names from training metadata, redirected artist-name prompts toward descriptive musical characteristics, and worked with Audible Magic and Musixmatch to screen uploaded audio and lyrics for potential unauthorized use. He also announced forthcoming download restrictions intended to curb mass distribution, and planned watermarking and fingerprinting technology.[1][5]

On November 25, 2025, Shulman announced Suno’s partnership with Warner Music Group, describing a path toward models built with licensed music and compensated experiences involving artists who opted in. His September 2026 v6 announcement broadened the stated audience to professionals and first-time creators, distinguishing models for precision, exploration and efficiency. He said previous models would be retired as v6 rolled out; further paid, opt-in experiences around individual artists remained in development.[4][6]