Dadabots

Dadabots represents a significant advancement in AI music creation, combining innovative technical approaches with a deep understanding of musical structure. By developing proprietary neural networks that generate raw audio directly, the company has pushed the boundaries of what AI can achieve in music production. This technical exploration stands at the intersection of computational musicology and artistic expression, offering valuable insights into the future of music creation. As AI technology continues to evolve, Dadabots' work provides a compelling case study in how these tools can complement human creativity rather than replace it. Through their development of specialized tools like DadaGP and GTR-CTRL, the company has created a technical foundation that enables both human-AI collaboration and independent AI-generated music, setting a precedent for future developments in the field.

Company Background and Technology Development

Dadabots' technology development began in 2012 at Music Hack Day MIT, where a team of independent musicians formed a hackathon team dedicated to creating remix bots that spidered SoundCloud for music to remix, posting hundreds of songs per hour. Their work drew inspiration from DeepMind and MILA's neural text-to-speech engines, specifically focusing on image style transfer techniques that demonstrated the ability to transform photographs into impressionist oil paintings.

The team undertook self-study in AI through resources like Tristan Jehan's "Creating Music By Listening" and the EchoNest Remix library, applying their learning to develop their proprietary system using deep learning techniques. They specifically experimented with SampleRNN and WaveNet technologies, which had previously demonstrated capabilities in human voice synthesis across multiple languages.

Their initial technical approach involved deep neural networks trained on short-term patterns (such as snare drum hits and scream timbres) and long-term patterns (guitar riffs and constant tempos), with diminishing returns observed during extended training phases. The company's technology primarily operates through an encoder/decoder tool that converts GuitarPro files into token sequences, drawing from a dataset containing 26,181 song scores across 739 musical genres.

The technical implementation requires significant computational resources, particularly V100 GPUs, due to the complexity of processing raw audio waveforms. This approach stands in contrast to most music+ML projects that generate MIDI outputs due to lower control requirements, while Dadabots' method produces audio directly, managing up to 10,000 times more data than sheet music per second. The team has trained hundreds of networks to find optimal hyperparameters, publishing their findings for broader use in the community.

AI Music Generation Process

Dadabots' AI music generation process begins with proprietary neural networks trained on raw metal album waveforms. These networks employ deep learning techniques specifically designed to capture short-term timbral patterns like snare drum hits and scream frequency modulation, alongside longer-term rhythmic structures such as guitar riffs and consistent tempos. The process demonstrates significant technical challenges, with the company noting diminishing returns in audio quality as training duration extends beyond initial phases.

The technical implementation requires sophisticated computational infrastructure, particularly NVIDIA V100 GPUs, due to the complexity of processing raw audio waveforms directly. This direct audio synthesis approach contrasts with most music generation work that produces MIDI outputs, highlighting the specialized requirements of working with raw audio signals. The current technical solution allows the system to generate audio at rates up to 10,000 times more data per second than comparable sheet music representations, though this intensive processing demands significant computational resources.

The output generation process involves several key steps. After network training on metal album waveforms, the system produces approximately 10 hours of raw audio output. This massive volume of data represents a substantial technical challenge, requiring careful curation by human operators to select the most viable musical segments for further development. The curatorial process typically takes about four days for album creation, demonstrating the time-intensive nature of transforming machine-generated audio into human-consumable music.

During the development process, the team has demonstrated multiple successful applications, including the generation of black metal and math rock compositions. Current technical capabilities allow the system to produce sustained audio outputs that maintain rhythmic coherence and musical structure, with some samples achieving a level of atmospheric quality well-suited for lo-fi black metal styles. The technical approach shows particular promise for complex genre fusion, allowing the system to combine multiple musical influences into a cohesive final product.

Technical Implementation and Data Processing

Dadabots developed a specialized set of tools and processing workflows to manage their complex AI music generation system. Their foundational technology includes DadaGP, an encoder/decoder tool that converts GuitarPro files into token sequences using a modified SampleRNN architecture. This dataset, containing 26,181 song scores across 739 musical genres, forms the basis for their neural synthesis work and enables human-AI collaboration in music production.

The company has developed several specialized tools to address the unique challenges of their workflow. DadaGP, their core file conversion tool, works with multi-instrument corpora including guitar, bass guitar, drums, piano, and orchestral parts. The system uses a tokenized format inspired by event-based MIDI encodings, making it suitable for generative sequence models while maintaining flexibility across multiple instrument types.

In addition to DadaGP, Dadabots developed ShredGP specifically for guitar tab generation. This tool uses a Transformer-based approach to imitate four distinct iconic electric guitarists, incorporating computational musicology methods to analyze DadaGP-encoded tokens. Two variants of ShredGP exist: one trained on a multi-instrument corpus and another on solo guitar data. The generated outputs are evaluated using a BERT-based model, demonstrating the system's capacity to produce guitar tablature consistent with specific player styles.

The company has also created GTR-CTRL, a genre-conditioned guitar tab synthesis tool that employs Transformer-XL architecture. This system incorporates special control tokens at the beginning of each song in the training corpus, allowing for precise manipulation of both instrumentation and genre-specific elements. The team has published results comparing model performance with and without these control tokens, demonstrating the tool's effectiveness in generating genre-specific guitar tabs while maintaining composition quality.

Dadabots' technical infrastructure includes D.O.M.E., a custom music exploration tool designed to manage large volumes of neural synthesis output. This system uses PCA-component k-means clustering with rasterfairy-quantized t-SNE grid visualization to navigate clusters of similar audio clips. Color mapping of spectral and chroma data enriches the visual representation, helping users intuitively understand sound relationships and range. The company's technical approach allows them to process raw audio waveforms at rates up to 10,000 times greater than sheet music representations, though this requires significant computational resources including NVIDIA V100 GPUs.

Applications and Collaborations

Dadabots has developed specialized tools for human-AI collaboration in music production, particularly for complex genres like black metal and math rock. Their work draws from mathematical principles and neural network architectures originally applied to image style transfer, demonstrating the potential for AI to generate music that transcends traditional compositional techniques.

The company's approach to music generation stands in contrast to most commercial music+AI projects, which focus on MIDI-based outputs due to lower control requirements. Dadabots' raw audio synthesis process requires advanced computational hardware, including V100 GPUs, to process the massive data volumes involved. This technical approach has produced several notable successes, including the generation of atmospheric black metal compositions that maintain rhythmic coherence across lengthy passages.

The team's work extends beyond technical development, incorporating elements of cultural exploration and artistic collaboration. Their research has demonstrated the potential of AI to reveal fundamental patterns in musical self-organization, even producing unexpected results like a neural network-generated scream about Jesus based on Kurt Cobain's acapella recordings. While the current process requires significant human curatorial intervention, the company views these challenges as essential stepping stones toward developing AI systems capable of independent musical creation.

Ethical and Creative Implications

Dadabots approaches AI music generation with a dual focus on technical innovation and ethical responsibility. While their technology has reached significant milestones, such as generating atmospheric black metal compositions and creating lo-fi black metal sounds, the team remains mindful of the broader implications of their work.

From an ethical standpoint, Dadabots views their technology as an extension of human musical capabilities rather than a replacement. They see AI-generated music as a form of cultural exploration, with efforts to maintain creative control through open research practices. The company's approach demonstrates a nuanced understanding of AI's potential impact, drawing parallels to effects pedals—while current audio quality may be subpar, the pursuit of high fidelity could lead to more centralized media control and disempower DIY artists.

The technical challenges of processing raw audio signals have led the team to advocate for open comprehension beyond open source, suggesting that music and art serve as ideal starting points for AI exploration. This stance reflects a broader awareness of AI's power dynamics, with the company noting that top AI companies are more powerful than most countries and are led by a small, knowledgeable elite.

In their applications of AI technology, Dadabots emphasizes the importance of human collaboration. The team's work with bands like Lightning Bolt and Krallice demonstrates their commitment to developing tools that assist music creators rather than replace them. Their technical infrastructure, including DadaGP for GuitarPro file conversion and D.O.M.E. for managing neural synthesis output, enables human operators to curate and refine machine-generated audio, highlighting their approach to AI as an extension of human creative processes rather than a replacement.