Source-linked AI summary
JazzSAMBA: A Synchronous and Asynchronous Multi-take Band Audio Dataset of Jazz Standards for Live Music Models
Phillip Long, Jacob Nguyen, Jace Hosto, Gage Hosto, Jett Takazawa, Fares Nofal, Sebastian Stade, Nithya Shikarpur, Julian McAuley, Cheng-Zhi Anna Huang, Stephen Brade, Aleksandra Teng Ma
TL;DR
Jazz research lacks clean, originally recorded multi-instrument combo stems of standards paired with structural annotations. JazzSAMBA provides such a dataset with asynchronous and synchronous takes, and its demonstrations show jazz-combo separation and chart-conditioned accompaniment evaluations where chart masks affect beat alignment more than coherence scores.
Problem
Jazz resources rarely provide originally recorded multi-instrument combo stems of standards, while accompaniment and form-aware analysis also need structural annotations aligned to those stems.
Source: §1 · p. 1Method
Results
Across both protocols, chart masks change Beat F1 more than COCOLA, while JazzSAMBA fine-tuning enables additional reportable separation stems and keeps BS-RoFormer within about 0.8 dB of its zero-shot result.
Source: §4.1 · p. 3, §4.2 · p. 4Takeaways & Limitations
Abstract
from arXiv · showhide
Machine learning has made strong progress on music tasks, both as assistive tools and as creative partners. However, most systems train on multitrack corpora that emphasize pop and rock. Jazz, with improvisation at the core of its practice, still lacks a well-annotated corpus of clean per-stem combo recordings on standards. We introduce JazzSAMBA (Jazz Synchronous and Asynchronous Multi-take Band Audio) to fill this gap: the first originally recorded jazz-combo multitrack dataset of standards with asynchronous (overdubbed) and synchronous (live ensemble) protocols, preferred and alternate takes chosen by the musicians, and timed annotations for bars, chords, sections, and soloists. JazzSAMBA covers 76 standards by eight musicians on drums, bass, piano, trumpet, and saxophone, with per-stem audio, mixtures, and MIDI. It can support chart-conditioned accompaniment, combo source separation, and form-aware music information retrieval. We demonstrate the dataset on two tasks: a jazz combo source-separation baseline and a chart-conditioned accompaniment ablation. The dataset, code, and samples are linked from the project demo page.
1. INTRODUCTION
Jazz research lacks clean, originally recorded multi-instrument combo stems on standards paired with performance-aligned harmony and form labels. JazzSAMBA introduces such a dataset, with two recording protocols and musician-selected takes.
1. INTRODUCTION
- Clean ensemble stems paired with structural annotations suit accompaniment, real-time generation, and source-separation research.Chart-conditioned generation and form-aware analysis can use the annotations as conditioning or evaluation targets.Source: §1 · p. 1
- JazzSAMBA provides clean per-stem combo recordings of standards with performance-aligned chord and section labels, addressing a gap in existing jazz resources.The dataset includes asynchronous overdubs and synchronous live recordings, with preferred and alternate musician-selected takes and timed bar and soloist annotations.Source: §1 · p. 1
2. RELATED WORK
JazzSAMBA addresses gaps in both general multitrack and jazz corpora by combining originally recorded jazz-combo stems with multi-take protocols and timed chart annotations. Its standards-based chart structure also aligns with interactive accompaniment research.
Multitrack datasets
- Unlike multitrack collections centered on pop, rock, or rendered MIDI, JazzSAMBA offers originally recorded jazz-combo stems of standards.The broader multitrack datasets support separation and transcription outside jazz.Source: §2 · p. 1
Jazz corpora
- JazzSAMBA targets the lack of originally recorded multi-instrument combo stems of jazz standards in prior jazz resources.Earlier corpora emphasized harmony and form, solo transcription, solo piano, or stems derived through source separation; H2H Music Improv covers free improvisation rather than a specific style.Source: §2 · p. 2
- JazzSAMBA unifies close-mic stems, dual recording protocols and takes, and timed chart annotations on standards.Its overview distinguishes overdubbed asynchronous recording from synchronous whole-ensemble recording; preferred mixtures receive timed labels, with alternate asynchronous labels copied and synchronous takes labeled separately.Source: §2 · p. 2, Figure 1 · p. 2
Interactive accompaniment
- Shared jazz charts couple improvisational freedom with form, a structure interactive accompaniment models can follow.Prior interactive systems have mostly used symbolic pop or classical music, or non-jazz audio.Source: §2 · p. 2
3. JAZZSAMBA DATASET
JazzSAMBA combines asynchronous overdubbing and synchronous ensemble recording, with take selection and annotation practices suited to each protocol. Its timed labels share a bar grid across a 76-standard corpus spanning five instrument categories and about 8.1 hours of preferred mixtures.
3. JAZZSAMBA DATASET
- Asynchronous sessions overdub instruments in sequence, while synchronous sessions record the full band together and are harder because of close-mic bleed.Each asynchronous instrument records two clean takes; preferred stems form one mixture and remaining takes form the alternate tier. Synchronous takes are annotated independently.Source: §3 · p. 2, Table 2 · p. 2
- Both protocols use a shared bar grid, with chords, sections, and soloists aligned to bar and fractional-bar offsets.Chord labels come from written lead sheets time-aligned by musicians, while section tags encode jazz forms and per-section tables record players and solo order.Source: §3 · p. 2
- The corpus covers 76 standards by eight musicians, with drums, bass, piano, trumpet, and saxophone across 50 asynchronous and 26 synchronous songs.Preferred mixtures total about 8.1 hours; both take tiers total about 16.1 hours, spanning 75–270 BPM.Source: §3 · p. 2, Figure 2 · p. 3
4. EXPERIMENTS
JazzSAMBA is evaluated on asynchronous combo source separation and chart-conditioned accompaniment trained across both recording protocols. The experiments show that jazz-specific fine-tuning enables additional reportable stems, while chart masks affect beat alignment more than coherence scores.
4.1. Source separation
- The separation experiment fine-tunes three models on asynchronous stem-sum mixtures for drums, bass, piano, and horns.It compares zero-shot, JazzSAMBA, and ChoraleBricks fine-tuning, evaluating full-track SI-SDR on five asynchronous test songs.Source: §4.1 · p. 3, Table 3 · p. 3
- Zero-shot BS-RoFormer is strongest on reportable stems, and JazzSAMBA fine-tuning keeps it within about 0.8 dB.JazzSAMBA fine-tuning makes horns reportable for HT-Demucs and piano and horns reportable for Open-Unmix, while also improving drums and bass on those models; ChoraleBricks fine-tuning does not improve JazzSAMBA test scores.Source: §4.1 · p. 3, Table 3 · p. 3
4.2. Chart-conditioned accompaniment
- The accompaniment ablation compares listen-only conditioning with section, chord, and combined chart masks using StreamGen.Models train on both protocols and are evaluated in 60-second free runs from form landmarks; scores include COCOLA and Beat Alignment F1.Source: §4.2 · pp. 3–4, Table 4 · p. 4
- Chart masks change Beat F1 more than COCOLA across both protocols, and the best mask often varies by stem.Adding chords often edges StreamGen's COCOLA over listen-only conditioning; synchronous Beat F1 is somewhat lower under live bleed, while COCOLA remains comparable to asynchronous results.Source: §4.2 · p. 4, Table 4 · p. 4
5. CONCLUSION
JazzSAMBA provides jazz-standard recordings with per-player stems and structural annotations, and the authors demonstrate source separation and chart-conditioned accompaniment uses. They hope it supports further work on interactive accompaniment and form-aware analysis.
5. CONCLUSION
- JazzSAMBA is the first jazz-standard dataset with per-player stems and annotations for chords, sections, and soloists, covering 76 standards performed by eight musicians.Source: §5 · p. 4
- The dataset supports music information retrieval and generation, and the authors demonstrate jazz-combo source separation and live chart-conditioned accompaniment benchmarks.Source: §5 · p. 4
- The authors hope JazzSAMBA will support interactive chart-following accompaniment, form-aware combo analysis, and models handling take-to-take and live-ensemble variation.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The recordings are original studio performances, and the dataset contains those recordings and the authors’ annotations.
6. COMPLIANCE WITH ETHICAL STANDARDS
- The musicians are credited as coauthors and consented to recording and public release; the dataset contains original studio performances and the authors’ annotations.Source: §6 · p. 4