Bar-level AI Composing Helper (BACH)

• Bar-level AI Composing Helper (BACH) is a system that generates human-editable, bar-aligned symbolic scores in ABC notation using a 'compose first, perform later' pipeline.
• It uses dual-stream tokenization and nested decoders to model synchronized vocal and accompaniment streams, ensuring enhanced musical controllability.
• BACH demonstrates superior performance in automatic and human evaluations, emphasizing practical improvements in controllability, perceptual quality, and song duration.

Searching arXiv for the core BACH paper and closely related interactive symbolic composition work. Bar-level AI Composing Helper (BACH) is a score-centric song-generation system built around the claim that the bar is the smallest meaningful semantic unit for songs because it encodes meter, strong and weak beats, and synchronization across melody, accompaniment, and multiple tracks. Rather than attempting to learn song generation directly from raw audio, BACH adopts a “compose first, perform later” pipeline: it generates a human-editable symbolic score in ABC notation and delegates final audio realization to downstream rendering tools. In this design, BACH itself generates symbolic scores, while the full system produces audio only after score rendering (Wang et al., 2 Aug 2025).

1. Historical position and problem framing

BACH is situated against a longer line of symbolic and interactive composition systems that treated controllability as a central problem rather than an afterthought. Earlier work on constrained symbolic composition combined human input with sequence models and finite-state constraints, extending Markovian constrained sampling to more expressive non-Markov models through sequential Monte Carlo and beam-search strategies for reharmonization (Walder et al., 2016). DeepBach then made Bach-chorale generation explicitly steerable, replacing left-to-right generation with conditionally editable note variables and pseudo-Gibbs sampling under positional constraints such as fixed notes, rhythms, cadences, metadata, and fixed voices (Hadjeres et al., 2016). The Bach Doodle demonstrated that short-span harmonization could be made interactive at scale by using Coconet in the browser, reducing harmonization latency from 40 seconds to 2 seconds for a two-measure task (Huang et al., 2019). Later interface work for pop-music infilling showed how per-track and per-bar controls, including track density, polyphony, occupation rate, and bar tonal tension, could be exposed in a musician-facing workflow over selected bars (Guo, 2022).

Within that lineage, BACH addresses a broader prompt-to-song setting and frames its contribution around four persistent limitations of prior song-generation systems: controllability, generalizability, perceptual quality, and duration. The paper argues that these problems arise because prevailing approaches try to learn music theory directly from raw audio, whereas BACH makes human-editable symbolic score generation the center of the system (Wang et al., 2 Aug 2025).

2. Compose-first, perform-later pipeline

The full BACH pipeline has three stages. First, an LLM, identified as Qwen3.0, converts a brief user request into multilingual lyrics and style metadata, organizing the lyrics hierarchically into sections such as intro, verse, chorus, bridge, outro. Second, BACH takes those lyrics and tags and generates a bar-aligned symbolic score in ABC notation. Third, the score is converted into performance audio: ABC is converted to MIDI, FluidSynth renders accompaniment audio, VOCALOID renders the sung vocal track, and the outputs are mixed into the final song (Wang et al., 2 Aug 2025).

This architecture is also the basis of BACH’s claim to human controllability. Because the intermediate object is a readable symbolic score rather than opaque audio tokens, users can edit the generated sheet music and lyrics directly, then re-render the song. The paper explicitly presents this as a remedy to black-box audio generation, where local changes typically require regenerating entangled audio containing composition, arrangement, timbre, performance nuance, and mixing. A common misconception is therefore that BACH is itself a raw-audio model. It is not: BACH is the middle symbolic composition stage in a larger score-to-performance system (Wang et al., 2 Aug 2025).

3. Symbolic representation and bar-level tokenization

BACH uses ABC notation, a text-based symbolic format that can represent pitch, rhythm, bar lines, lyrics, and multi-voice information. The representation is operationalized through bar-stream patching rather than byte-level or BPE tokenization. A score is first segmented by bar; each bar is then split into non-overlapping 16-character patches; and if the final patch is shorter than 16 characters, it is zero-padded. The vocabulary is defined as all possible 16-character ASCII strings, plus three special symbols; the method description also uses control markers such as \langle [SOA](https://www.emergentmind.com/topics/smooth-offset-algorithm-soa) \rangle, \langle EOA \rangle, \langle EOD \rangle, [[START](https://www.emergentmind.com/topics/sub-task-aware-robotic-transformer-start)], and [[END](https://www.emergentmind.com/topics/eccentric-nuclear-disk-end)] (Wang et al., 2 Aug 2025).

This tokenization reflects the paper’s core representational thesis. Timing inside a bar is not recoded into a separate event vocabulary; instead, meter, note lengths, rests, ties, bar lines, and lyric alignment remain encoded through the original ABC characters. Track structure is handled through Dual-NTP, which routes each patch token into one of two streams: vocal and accompaniment. The representation therefore makes three commitments simultaneously: the bar is the primary semantic unit, score text remains human-readable, and vocal and accompaniment streams are modeled as distinct but synchronized symbolic channels (Wang et al., 2 Aug 2025).

A plausible implication is that BACH’s notion of “bar-level” is not merely user-interface terminology. It is built into tokenization, timing organization, and the separation between high-level composition and downstream performance rendering.

4. Hierarchical generation procedure and formal structure

BACH combines two hierarchies: song-section hierarchy and bar-level symbolic hierarchy. The main structural conditioning mechanism is Chain-of-Score (CoS), which serializes instruction, style tags, raw lyrics, and section-wise score spans into a single training document:

$$\mathcal{D}_{\text{cos}} = \text{Instruct} \circ \text{Tag} \circ \text{Lyrics} \circ \left( \bigcirc_{i=1}^{N} s_i \right) \circ \langle EOD \rangle$$

Each section is serialized as

$$s_i = [START] \circ \tau_i \circ \ell_i \circ \langle SOA \rangle \circ \psi_i \circ \langle EOA \rangle \circ [END]$$

where $\tau_i$ is the structure label for section $i$, $\ell_i$ is the section lyric content, and $\psi_i$ is the sequence of Dual-NTP score tokens (Wang et al., 2 Aug 2025).

Architecturally, the score generator uses two nested decoders. A patch-level decoder models dependencies between score patches, while a character-level decoder autoregressively predicts the characters of the next patch. The paper also recalls the standard next-token factorization

$$p(\mathbf{x}_{1:T}) = \prod_{t=1}^{T} p(x_t \mid x_{<t}; \theta)$$

and specializes it to Dual-NTP by jointly predicting vocal and accompaniment streams:

$$p(\mathbf{v}_{1:T}, \mathbf{a}_{1:T}) = \prod_{t=1}^{T} p(v_t, a_t \mid v_{<t}, a_{<t}; \theta)$$

with inference written as

$$(\hat{v}_t, \hat{a}_t) = \arg\max_{(v_t, a_t)} p(v_t, a_t \mid v_{<t}, a_{<t}; \theta)$$

The paper also defines in-context style transfer by prepending a reference excerpt:

$$\mathcal{D}_{\text{icl}} = \mathcal{A}_{\text{ref}} \circ \mathcal{D}_{\text{cos}}$$

(Wang et al., 2 Aug 2025)

Training uses a large symbolic corpus described in token counts rather than song counts: 1B conditional vocal tokens, about 10B unconditional music tokens, and 2B CoS music tokens. During annealing, 1B CoS tokens from a high-quality subset are replicated fourfold to create a 4B ICL dataset. The reported mixture ratios are conditional : unconditional = 1 : 3 and music : speech = 10 : 1 before annealing, then CoS : ICL = 2 : 1 during annealing. The evaluated system is BACH-1B, indicating roughly a 1B-parameter model. Test-time decoding includes forced decoding inside the score span, top-k = 50, top-p = 0.93, temperature = 1, repetition penalty = 1.1, max new tokens = 3000, and classifier-free guidance with scale $$s_i = [START] \circ \tau_i \circ \ell_i \circ \langle SOA \rangle \circ \psi_i \circ \langle EOA \rangle \circ [END]$$0 (Wang et al., 2 Aug 2025).

5. Human controllability and symbolic editing

BACH’s controllability operates at several levels. At the prompt level, the user specifies style and content. At the lyrics and style-tag level, Qwen3.0 expands the prompt into multilingual lyrics and explicit section annotations. At the structural level, Chain-of-Score makes sections such as intro, verse, chorus, bridge, and outro explicit. At the bar level, users can modify melody, rhythm, lyric-to-melody alignment, accompaniment content, and local phrase structure directly in the symbolic score. At the track level, Dual-NTP separates vocal and accompaniment streams so that one can be revised without necessarily disturbing the other (Wang et al., 2 Aug 2025).

This symbolic editability distinguishes BACH from raw-audio systems and also from earlier Bach-specific interactive models. DeepBach, for example, achieved steerability by making every note a conditionally editable variable in a four-part chorale representation; BACH instead generalizes the idea to prompt-conditioned long-song generation with lyrics, section structure, vocal/accompaniment separation, and symbolic score rendering (Hadjeres et al., 2016). The result is not merely “control” in the sense of prompt steering, but an explicit human-in-the-loop editing substrate.

The same point clarifies another common misconception. BACH is not a hidden-score model that only exposes audio afterward. The score is the principal artifact. The paper repeatedly presents this as the reason users can preserve satisfactory parts, revise unsatisfactory bars, alter local rhythm or melody, and re-render without resampling an entire song (Wang et al., 2 Aug 2025).

6. Empirical results, limitations, and broader significance

The evaluation protocol compares BACH-1B against YuE, YuE-light, Suno, Udio, Hailuo, Tiangong, and partially SongComposer, with each model generating 20 full-length English songs from diverse prompts. Reported automatic metrics include KL Divergence, FAD, CE, CU, PC, PQ, CLAP, and CLaMP 3. In the table reproduced in the paper, BACH attains KL 0.391, FAD 1.526, CE 7.323, CU 7.976, PC 6.531, PQ 8.220, CLAP 0.212, CLaMP 3 0.263, and the best reported overall score of 28.608. The paper states that BACH sets SOTA on three individual metrics and the overall composite, ties the top PQ score at 8.220, and beats commercial systems on CLaMP 3 alignment with 0.263 (Wang et al., 2 Aug 2025).

Human evaluation uses 50 raters—10 AI researchers, 5 trained musicians, and 35 non-experts—in blind A/B comparisons. The detailed discussion reports that BACH significantly outperforms Hailuo, Tiangong, and YuE-light, moderately surpasses YuE and Udio, and is competitive with Suno, though still trailing Suno overall in musicality. It is described as especially strong in vocal quality, melodic attractiveness, accompaniment quality improvements over open models, and vocal-backtrack matching. The paper also states that BACH produces the longest average duration and greatest upper-bound song length among compared systems, and claims minute-level generation versus hours for the strongest open-source baseline (Wang et al., 2 Aug 2025).

The limitations are explicit. Final audio quality is still bounded by the renderer, since accompaniment is rendered with FluidSynth and vocals with VOCALOID. The system does not use a dedicated learned score-to-audio model. Open high-quality song datasets remain limited. Automatic metrics are acknowledged as imperfect proxies for human perception. In controllability, the system is reported to be especially good at tempo/rhythm control and instrument/vocal control, but weaker at genre control and emotion control (Wang et al., 2 Aug 2025).

In a broader symbolic-composition context, BACH’s emphasis on editable notation aligns naturally with symbolic evaluation methods. Interpretable grading functions for four-part Bach-style chorales, based on pitch, rhythm, per-voice intervals, harmonic qualities, parallel errors, and repeated sequences, have already been proposed as automatic critics for symbolic generation (Fang et al., 2020). This suggests that score-centric generation and score-centric evaluation are technically complementary. A plausible implication is that BACH’s most durable contribution is not only a set of leaderboard results, but a reframing of song generation itself: from “generate audio that contains music” to “generate music, then perform it.”