Creating Music with AI: State of the Art and Perspectives

Executive Summary

Music AI today encompasses powerful technologies for generating entire songs or instrumental tracks, based on deep models (neural networks, Transformers, diffusion, GANs) in both symbolic domain (MIDI) and raw audio[1][2]. After a technical overview and description of the complete pipeline (preprocessing, training, generation, post-processing), we present recent examples (OpenAI Jukebox, Google MusicLM, Meta MusicGen, Moûsai diffusion, Google Magenta, as well as emerging Suno/Udio/Soundraw, etc.). A detailed comparison table gathers the main platforms (Jukebox, MusicLM, Magenta, AIVA, Amper, Soundful, Splice, LANDR, Stable Audio, Mubert, Boomy, Suno, Udio, Soundraw…) by their features, supported formats, audio quality, creative control, pricing/licenses, interoperability and customer reviews. We analyze the developments of the past 12–24 months (technical advances, open vs closed models, emerging trends) and evaluate advantages (accelerated creativity, accessibility), drawbacks and risks (cultural biases, uneven quality, rights issues, employment impacts). Finally, we provide practical recommendations for musicians, labels and startups (tool selection according to workflow, legal questions to clarify, best practices).

Section 1: Technical Principles and Definitions

Generative music AI consists of using machine learning models to automatically create music. Approaches are divided into two broad categories:

Symbolic domain (MIDI/scores) vs raw audio: Symbolic models generate sequences of notes/instruments (piano-roll, MIDI files). Examples: Google Magenta (Music Transformer[2]), OpenAI MuseNet, Magenta's PerformanceRNN, which excelled in recent years at composing polyphonic classical or pop works. They capture musical structure (melody, rhythm, harmony) but do not directly produce sound (no timbres or voices). In contrast, audio models operate on the signal itself (waveforms, spectrograms). They can generate complete tracks with timbres and voices, but require enormous computing capacity to model the wave at high fidelity. This is the case with Jukebox (OpenAI)[1] or MusicLM (Google)[3].

Model architectures:

• Transformers/autoregressive: Many tools (Jukebox[1], MusicGen, MusicLM[3]) use Transformers to predict the next audio sequence or musical token. For example, Jukebox first encodes audio into discrete codes (VQ-VAE) then performs a hierarchical Transformer on these codes to generate the music[1]. Music Transformers (Magenta) apply to MIDI and improve long-term coherence[2].
• Diffusion: Inspired by GANs, these models learn to gradually remove noise to generate audio. For example, Moûsai (ACL 2024) is a two-stage diffusion model generating several minutes of high-quality stereo music (48 kHz) from a text prompt[4]. Stable Audio (Stability AI) also uses latent diffusion to create complete tracks (version 2.0 reaches 3 min at 44.1 kHz)[5].
• GANs and others: Although less common recently for long works, GANs (e.g. MuseGAN) or RNN/LSTM networks have been tested to generate musical bars. They are now often surpassed by the fluidity of Transformers/diffusion on complex content.
• Hybrid models: Some pipelines combine symbolic steps (melody opening) and audio synthesis. For example, a model can first compose a MIDI (Music Transformer) then use another network (WaveNet) to synthesize the final audio[6].

In summary, current models are distinguished by their ability to generate either symbolic representations (notes/MIDI)[7] or direct audio signals[1][3]. They use Transformers to model musical temporality over long durations, and diffusion to improve the final sound quality on complete tracks.

Section 2: Complete Music Generation Pipeline

AI music creation generally follows these steps:

1. Data collection and preprocessing: A large musical corpus (audio and/or MIDI) is created or retrieved. This corpus is cleaned and formatted: track separation (bass, vocals, etc.), structural analysis (verses, choruses) and annotation of musical metadata. For example, the SongPrep chain introduced in 2024 automatically separates sources, identifies the song's structure and even transcribes lyrics, thus producing structured musical data for training[8]. Audio is often converted into intermediate representations (spectrograms, VQ-VAE codes, embeddings) to facilitate learning.
2. Model training: One or more models are trained on this data. If working in audio, one can train an autoencoder (like a VQ-VAE) to compress audio into discrete codes, then a generative model (Transformer, diffusion) to learn to reconstruct these codes. For example, Jukebox uses a hierarchical VQ-VAE then 3 autoregressive Transformers (one per code level)[1]. In the symbolic case, a sequence model (RNN/Transformer) is directly trained on musical events (notes, durations). Some industrial tools (MusicLM) train hierarchical sequences coupled with text[3]. The role of this phase is to incorporate into the model the statistical laws of music (rhythm, chords, timbres) present in the corpus.
3. Generation / inference: The model generates new music by sampling from its learned distribution. For an autoregressive model, a seed (textual or musical prompt) is initialized and the audio or symbolic sequence is progressively generated. For example, Jukebox generates musical codes level by level (from the most abstract level to timbral detail) then decodes to audio[9]. A diffusion model starts from random noise and refines the spectrogram until obtaining the final audio. In practice, generation is often guided by a text prompt ("a soft pop ballad with female vocals") or by a short musical sample to complete. Controllable parameters can include style, tempo, duration, key, and even instructions on instruments.
4. Post-processing: The raw result is then processed to become a usable final product. Automatic mastering can be applied (e.g. LANDR offers full AI mastering of generated tracks), effects added (compression, reverb), volume normalized, or stems separated. A DAW can also be used to edit the structure (cut/paste, arrange). Depending on the tool, the final output consists of audio files (MP3/WAV) or directly separated stems. Some automated workflows also propose adding lyrics or synthetic singing afterwards, thanks to separate vocal models (for example, synchronizing a voice model on a generated melody). The goal of post-processing is to ensure audio coherence (no clicks, no tempo jumps) and adapt the composition to needs (mp3 format, loop, separation for remix, etc.).

Schematically, as illustrated by the SongPrep chain for songs, one goes from raw data (audio/MIDI) to structured data for training (track separation, structure, lyrics)[8], then generates new tracks that are refined in the final step. Each step relies on varied algorithms: source separation, neural networks for generation, then classic DSP in post-processing.

Section 3: Recent Architectures and Algorithm Examples

Here are some notable achievements and their technologies:

• OpenAI Jukebox (2020)[1][9]: generative model of raw audio. It first compresses audio into three levels of codes (hierarchical VQ-VAE), then trains three autoregressive Transformers to model the code sequence. The top level is generated first, then the other levels are "resampled," before decoding to audio. Jukebox produces songs sometimes including singing, but with quality still distant from the studio (vocal textures somewhat "robotic").
• Google MusicLM (2023)[3]: text→music generator. MusicLM uses a hierarchical seq2seq model operating on mel-spectrograms. It can create several minutes of coherent music at 24 kHz from a textual description. According to Google, MusicLM surpasses in quality and text fidelity all previous music generation systems[3]. It also integrates a "melody-to-music" mode allowing a melody given by a user to be refined, and an "unconditional" mode for instrumental patterns. This system remains a research prototype, not (yet) commercial.
• Meta MusicGen (2023)[10]: unique diffusion/transformer model (Audiocraft) that directly generates stereo samples (mono and stereo) at 24 kHz. It takes text or a melody as input, and produces the music in a single pass. MusicGen has the advantage of being relatively simple and modular, able to specialize in several versions (instrumental, male/female vocalist). The rendering is high quality and faster to generate than with Jukebox.
• Hierarchical diffusion (2024): Moûsai (Peking University, 2024) is an open-source two-stage latent diffusion model that creates stereo audio tracks of several minutes (48 kHz) from a text prompt[4]. It is designed to be very efficient (real-time inference on a medium GPU). Other experimental works use complex diffusion chains (e.g. Meta's Splash diffusion or Stable Audio). Stable Audio 2.0 (Stability AI, 2026) uses a transformer diffusion to generate tracks up to 3 minutes at 44.1 kHz, including an audio-to-audio transformation mode[5].
• Google Magenta and other R&D: Magenta remains active in academic R&D. For example, Google Magenta's Music Transformer (2018) uses relative attention to generate polyphonic MIDI sequences that are very coherent over the long term[2]. Magenta also explores VAEs (MusicVAE) and style transfer models. In parallel, companies such as Suno, Udio or Soundraw develop their own unpublished models, sometimes in collaboration with university research.
• Suno AI (2023-2026): startup from Cambridge (UK) founded in 2023[11], now based in San Francisco[11], Suno offers a web interface for generating short music (pop, dance, etc.) from text[12]. Their version 5.5 introduced voice capture in generation[13]. They made the news with a lawsuit with the majors (RIAA) for the use of protected data[14], settled in 2025 by an agreement requiring the use of a licensed dataset[15].
• Udio AI (2026): emerging platform (site udoi.com), presented by Soundverse AI as a complete tool for generating complete songs from text prompts, melodies, or audio references[16][17]. Udio stands out for several advanced modes ("Describe Your Song," "Custom Mode," support for reference files) and multilingual support. According to Soundverse, it handles lyrics, melody composition and mixing entirely[16][17].
• Soundraw (2018–…): Japanese AI music platform oriented towards content creators. Soundraw allows generating custom instrumental tracks (genre, mood, instruments)[18]. Its official blog highlights ethical AI (not trained on protected content[19]) and a flexible licensing model (track commercially usable even after unsubscribing[20]). Soundraw particularly targets the needs of video/marketing/podcast, where original music is desired without worrying about copyright.

Each architecture differs in possible inputs (text, audio files, starting melodies or rhythms) and in output (mono/stereo tracks, with or without vocals)[16][12]. Recent models prioritize modularity (Text-to-music + Voice-to-music + Instrumental) and high-fidelity audio quality.

Section 4: Existing Solutions — Detailed Comparison

We compare here the main AI music platforms, with a focus on Suno, Udio, Soundraw (among others). The table below synthesizes features, formats, quality, creative control, pricing/license, interoperability, use cases, customer reviews and limitations for each solution. Information is drawn from official sources (websites, technical blogs) and user feedback (Trustpilot, forums)[12][17][21].

The table is based on official websites and user feedback. For example, Suno highlights a free generator without credit card, producing "studio-quality" tracks downloadable in high definition[12][44]. Udio boasts a complete creation pipeline (lyrics→melody→mixing) with advanced customization modes[16][17]. Soundraw insists on its ethical constitution (AI not trained on protected works[19]) and the ability to freely download tracks even without a subscription[42].

Customer reviews extracted from Trustpilot highlight practical flaws: several users describe Suno as "unusable" due to bugs (sudden speed-ups)[38], Udio is judged qualitatively solid but criticized for terms of use "too restrictive"[41], and Soundraw is criticized for generating repetitive tracks ("absolute garbage, use Suno or Udio")[43]. Conversely, established solutions like LANDR benefit from generally positive reviews (Trustpilot 4.0/5)[33] thanks to the quality of their AI mastering.

Comparison Table

Section 5: Recent Developments (2024–2026)

Since 2024, music AI has made rapid progress:

Timeline (2023–2026) of Music AI Advances

• 2023: Launch of Suno AI (Cambridge, announcement of an improved "v4.5" model)[11]
• 2023: Commercial launch of several tools (Boomy, Soundraw, etc.)
• 2024: Release of Meta's MusicGen model (NeurIPS 2023)[10]
• 2024: Moûsai (diffusion) published, capable of generating several minutes of 48kHz music[4]
• 2024: Suno sued by RIAA for use of protected music[14]
• 2025: Suno/WMG (major) agreement for models trained on licensed corpus[15]
• 2025: Emergence of Udio AI (complete lyrics→music platform) and integration into pro workflows[16][17]
• 2026: Launch of Stable Audio 2.0 by Stability AI (text-to-audio tracks 3min @44kHz)[5]
• 2026: Suno opens a San Francisco office to expand its ML teams[11]

Recent trends include:

• Rise of open source models: Beyond the giants, many startups and labs publish their weights (diffusion / Transformers) on HuggingFace and GitHub. For example, Meta's MusicGen, OpenAI's VQ-Diffuser, or academic projects like Riffusion and Moûsai make new tools accessible, forcing commercial solutions to innovate further.
• Increased controllability: Systems now integrate more controls. Interactive interfaces are proliferating (e.g. Google MusicFX DJ for real-time mixing[45]) or advanced options (Udio AI allows audio references, Soundraw adjusts instruments and tempo). The goal is to give users more control (exact tempo, section structure, vocals or not, etc.).
• Audio quality and duration: Quality has improved (stereo audio, realistic voices) and generated durations are lengthening (very long pieces thanks to hierarchical architecture or diffusion). Audio resolutions are rising (from 16 kHz with early prototypes, we are now at 24–48 kHz for MusicLM, Moûsai or Stable Audio).
• Closed vs open models: Some solutions remain proprietary (MusicLM, Udio, AIVA), others are open-source (Magenta, Musenet, Moûsai, MusicGen). In 2025-2026 a hybridization is noted: startups like Suno and Soundverse partially publish their advances or datasets to reassure on legality, while players like Stability share their models on an open platform (Stability AI).
• Regulation and licenses: Following lawsuits, the emphasis is on dataset ethics. Initiatives like Fairly Trained or AI:OK are developing to certify legal data. For example, Soundraw publicly states it uses no protected content[19], and Suno had to commit to training its AI on licensed content[15].

Overall, the ecosystem has entered a maturity phase, moving from simple loops (2021-2022) to complete musical generation viable for professional uses (interactive demos, DAW integration, AI plugins). The timeline above illustrates some key milestones, notably the rapid evolution from 2025 towards more robust and ethical tools.

Section 6: Advantages, Limitations and Risks

Advantages

Music AI offers a powerful extension of creative capabilities:

• Time savings and enhanced creativity: generating a composition or a draft in a few seconds, quickly moving from concept to sound experience. Musicians can thus experiment with new themes or provide instant demos at lower cost.
• Democratization of music: non-musicians can compose custom soundtracks (for videos, podcasts, advertising) without knowing how to play an instrument. This opens up markets (content marketing, jingles) and allows small productions to compete.
• New economic opportunities: emergence of AI SaaS models (subscriptions, APIs) and licensing opportunities (for example, Suno launches industrial contracts after its lawsuit[15]). Some labels create their own AIs (e.g. Capitol AI) to offer music "in the style of" major catalogs.
• Efficient workflow: integration into existing production tools (plugins, DAW assistants, AI mastering) simplifies the studio pipeline. For example, LANDR or Studio.ai add instant mastering, Soundraw or Splice instrument plugin add AI composition as a complement.

Limitations and Risks

However, several problems persist:

• Variable quality: outputs are not always robust: repetition problems, strange harmonics, synthetic voices still perceptible. Several users note that the AI "often goes haywire" or "speeds up for no reason"[38][43]. Current models reach studio quality for short-term or simple genres, but struggle with complex structures (tempo variations, subtle emotions).
• Limited artistic control: although platforms add options, generation remains largely random. Changing a detail (one more beat, replacing an instrument) is not trivial: it generally requires regenerating an entirely new track. Professional artists may fear a lack of control over the final work.
• Cultural and style biases: models reproduce the majority trends of their dataset (e.g. strong preponderance of Western pop/rock works). Less represented genres (traditional music, complex jazz) are often less well handled. Tempo, instrumentation, or even gender biases are noted (e.g. AI generating male/very feminine voices in some cases).
• Copyright and ethics: training on billions of excerpts raises serious licensing questions. Labels (RIAA) have pursued startups like Suno accusing "large-scale theft"[14], forcing agreements (WMG/Suno) framing training. The subsequent use of creations raises questions: who is the legal author of a track 100% generated by machine? Several platforms like LANDR or YouTube now require declaring AI use and reserve the right to block entirely generated works[46]. Musicians must ask the right legal questions (origin of training data, rights granted, possible distribution).
• Employment impact: the fear of partially replacing sound engineers, jingle composers or session musicians is real. Some thinkers point out that AI could destroy low-level creative jobs (roles that are "very repetitive"), and even affect artists' rights for the benefit of technology players. However, others believe that AI frees up creative time to refine the art, while creating new professions (prompt engineering, data curators).

Section 7: Impacts on the Music Industry

The integration of AI is disrupting several areas of the industry:

• Artistic creation: massive increase in content produced. Platforms distribute millions of tracks per month (Boomy claims hundreds of thousands of users). Illustration music (video, games, advertising) is seen saturated with generated content. At the same time, songwriters seek to integrate AI into their process (e.g. co-writing with AI, AI remix). Freemium models allow small labels or independent creators to produce without budget.
• Technical production: studios and DAWs integrate AI assistants (e.g. auto-mastering, track separation, melody completion). Production costs drop (no longer needing equipment or engineers for certain tasks). Software chains (plugins, cloud platforms) multiply to accommodate AI in effects/sound stores.
• Distribution: music aggregators (Spotify, iTunes) adapt their policies. Some now require metadata on AI use, or block 100% AI works in certain playlists. Business models pivot towards subscription: AI becomes a subscribed service (even streaming platforms are considering a "Netflix model for AI music"). Some companies promote "on-demand music generation" with shared royalties.
• Rights and collection: new mechanisms for tracking musical provenance are emerging. For example, startups are developing technologies to detect whether a track was generated by AI or contains copied elements[47]. Collective management organizations are reviewing how to distribute rights on works using AI samples. The LANDR contract clearly states that any 100% AI track may be considered as not falling under standard distribution[46], illustrating the legal complexity.
• New business models: In addition to subscriptions, dedicated marketplaces are appearing (e.g. Songer for selling AI tracks) and music token ICOs (unverified but mentioned in crypto press). Music AI also creates synergies: e.g. video production companies integrating end-to-end AI audio/video, and even NFT "virtual musicians" commercializing their AI "creations." Investments in music AI explode at the end of 2023-early 2025, even if the financial climate remains uncertain (succession of funding rounds and lawsuits).

Section 8: Synthesis of Customer Reviews

User feedback (French and English-speaking) illustrates the practical strengths and weaknesses of tools. Here are some representative excerpts from platforms like Trustpilot and specialized forums:

• AIVA: users disappointed by complicated editing. Example: "I took the Pro subscription, but after 30 minutes the platform didn't deliver on its promises… the system doesn't follow instructions."[25]. Some regret the average quality for professional production.
• Splice: great success for its audio library. A user notes: "Excellent service. But the credit policy is very bad… you lose credits when unsubscribing, pushing consumption"[31]. In summary, the sound richness is praised, but the credit-based business model is criticized.
• Soundful: very negative reviews on Trustpilot. For example: "This site is a total scam… they charged my account without warning, no way to cancel, run"[28][48]. Customers complain about bad commercial practices (unauthorized charges) more than the product itself.
• Boomy: disastrous rating (Trustscore ~1.8). A musician reports: "Boomy is a scam: I had 2.9M streams and they didn't pay me a cent…"[36]. Many testimonials denounce account freezing and non-payment of streaming revenues. The revenue sharing model seems to malfunction.
• Mubert: also very poorly rated (1.7). Extract: "Horrible site. I was scammed out of nearly $1,000. No refunds. Non-existent customer service"[35]. Reviews highlight a disappointing and potentially fraudulent user experience.
• LANDR: on the contrary, praised for its services. A user writes: "Pro-level mastering tool, very easy to use. Efficient distribution without issues."[49]. However, critics point to slow assistance or minor rights verification issues.
• Suno AI: new criticisms (Trustpilot 1.7/5). For example: "Unusable. Tracks speed up in the middle… from 100 BPM to 250 BPM for no reason[38]." Another notes: "Lots of potential, but the AI often goes haywire…"[50]. The community therefore reports bugs and limited prompt compliance.
• Udio AI: mixed reviews. An expert confirms the audio quality but objects: "Impressive audio quality, but terms of use too restrictive for serious use"[41]. Some English-speaking users report similar experiences regarding licenses.
• Soundraw: also widely criticized (2.0/5). A user writes bluntly: "Absolutely rubbish music, the AI always generates the same song… use Suno or Udio instead"[43]. Another found the tool "rudimentary and weak" despite promises of perpetual licenses[51]. Only a few testimonials praise its ability to quickly produce jingles without costly licensing, but they are in the minority.

Synthesis: established and older solutions (Splice, LANDR) generally receive good reviews in their main domain, while AI newcomers (Suno, Udio, Soundraw, Mubert, Boomy) are strongly divisive. Common criticisms concern the artistic quality of the result (repetitiveness, voice intelligibility, interface bugs) and commercial reliability (cancellations, refunds, late payments). Users often advise testing several tools and not relying solely on commercial testimonials.

Section 9: Practical Recommendations

For musicians, labels and startups wishing to leverage music AI, here are some practical tips:

• Choose the right tool for your project: If you need simple loops or jingles, Splice, Soundraw or Amper may suffice. For more complex tracks (melodies, lyrics), consider AIVA, Suno or Udio. For final mastering/refinement, LANDR or iZotope Ozone (AI) are appropriate. Evaluate according to the desired format (separate stems, mp3, midi) and the level of control required. API solutions (Amper, Stable Audio) are suited to software integrations, while web interfaces (Suno, Boomy) are very accessible but less customizable.
• Check legal aspects: Always explicitly ask under what license the generated music is provided. For example, Suno and Soundraw announce "100% royalty-free" content[12][21]. But terms of use can exclude certain uses (or require a Pro subscription to transfer rights[52]). Ask the provider about the origin of the training dataset (is it entirely legitimate? Has the company been sued, like Suno with the labels[14]?). In many countries, jurisprudence on this subject is still being defined: remain cautious about commercial exploitation and consider crediting the tool if necessary.
• Integrate AI into the workflow: Use AI as a co-creator, not as a turnkey solution. For example, use Suno/Mubert to generate rhythm ideas, then import them into your DAW to rework. Or use Udio to generate a first melody, then stop there and rebuild the structure yourself. In a professional context, combine AI and human skill: a composer can refine an AI skeleton by adjusting the final mix and adding live elements. Test several prompts and refinements to not limit yourself to a single result.
• Stay informed and iterate: The field evolving rapidly, try new versions (Suno 5.x, MusicGen 2.0, etc.) as soon as they are released. Join communities (Splice Discord, ML forums) and subscribe to tech newsletters (Water & Music, Axios). This allows knowing tips (e.g. how to formulate an effective prompt) and problems (e.g. timing offset on certain platforms).
• Quality management: Always carefully listen to the AI production. Check tempo consistency and absence of glitches (as reported for Suno[38]). If targeting radio or streaming, systematically go through professional mastering (human or AI) to ensure optimal sound level and clarity. Think about balancing AI originality with a personal touch (re-recording a solo, for example) to differentiate your works.
• Economic aspects: Budget unknown, but evaluate ROI. Many platforms offer free trials; take advantage to compare. Note that pro plans (Amper, Splice Creator) can be expensive (≈€10–40/month), but often offer more control and rights. 100% free solutions (AIVA Pro transfers rights[52], Soundraw without DRM[42]) are preferable if targeting intensive commercial exploitation. Also calculate revenue sharing (Boomy keeps half, for example).

Conclusion

Music AI is now establishing itself as a new essential tool for audio professionals, seeking agility and production volume. Recent advances (audio quality, granular control, advanced integrations) make the technology attractive, but caution remains necessary given the ethical and legal stakes. Our comparative analyses and recommendations aim to guide informed users: whatever the project (quick composition, demonstration, final track or mastering), there is now a specialized tool. However, none replaces human creativity: AI must be handled like an instrument (both powerful and imperfect).

Sources and References

• [1] [7] [9] Jukebox | OpenAI — https://openai.com/index/jukebox/
• [2] [6] Music Transformer: Generating Music with Long-Term Structure — https://magenta.withgoogle.com/music-transformer
• [3] MusicLM — https://google-research.github.io/seanet/musiclm/examples/
• [4] Moûsai: Efficient Text-to-Music Diffusion Models - ACL Anthology — https://aclanthology.org/2024.acl-long.437/
• [5] [34] Introducing Stable Audio 2.0 — Stability AI — https://stability.ai/news-updates/stable-audio-2-0
• [8] SongPrep: A Preprocessing Framework — https://arxiv.org/html/2509.17404v1
• [10] Simple and Controllable Music Generation — https://arxiv.org/abs/2306.05284
• [11] [14] [15] [39] Scoop: AI music generator Suno opens SF office - Axios San Francisco — https://www.axios.com/local/san-francisco/2026/04/03/suno-ai-music-generator-startup-office-expansion
• [12] [37] [44] Suno AI Music: Free AI Music Generator — https://sunoai-music.com/
• [13] Suno launches v5.5 AI model with voice cloning tool — https://www.musicbusinessworldwide.com/suno-launches-v5-5-ai-model-with-voice-capture-and-personalization-features/
• [16] [17] [40] What Is Udio AI? — https://www.soundverse.ai/blog/article/what-is-udio-ai-0139
• [18] [19] [20] [21] [42] SOUNDRAW Blog — https://soundraw.io/blog/post/introduction-to-soundraw
• [22] [23] [24] [52] AIVA — https://www.aiva.ai/
• [25] Aiva Reviews Trustpilot — https://www.trustpilot.com/review/www.aiva.ai
• [26] [27] Amper Music — https://www.toolhunter.ai/ai-tool/amper-music
• [28] [48] Soundful Reviews Trustpilot — https://www.trustpilot.com/review/soundful.com
• [29] [30] Splice AI — https://www.toolcentral.ai/ai-tools/splice-2/
• [31] Splice Reviews Trustpilot — https://www.trustpilot.com/review/splice.com
• [32] LANDR — https://www.landr.com/
• [33] [49] LANDR Reviews Trustpilot — https://www.trustpilot.com/review/landr.com
• [35] Mubert Reviews Trustpilot — https://www.trustpilot.com/review/mubert.com
• [36] boomy.com Reviews Trustpilot — https://www.trustpilot.com/review/boomy.com
• [38] [50] Suno Reviews Trustpilot — https://www.trustpilot.com/review/www.suno.ai
• [41] Udio Reviews Trustpilot — https://www.trustpilot.com/review/udio.com
• [43] [51] SOUNDRAW Reviews Trustpilot — https://www.trustpilot.com/review/soundraw.io
• [45] [47] Three music AI trends to watch in 2025 — https://www.waterandmusic.com/three-music-ai-trends-to-watch-in-2025-recording-takeaways/
• [46] AI music tools and music distribution | LANDR — https://www.landr.com/ai-music-tools-and-music-distribution