This document presents a method for generating walking bass lines using hidden Markov models. It proposes three methods for defining hidden states and evaluates them objectively and subjectively. Method 3, which considers pitch class and metrical position, performed best by generating bass lines that were harmonically congruent, sequentially smooth, and preferred by an expert bassist over the ground truth. While effective, the approach has limitations like only handling short, simple chord progressions and more work is needed to address longer, more complex progressions and additional musical factors.
JamSketch: Improvisation Support System with GA-based Melody Creation from Us...kthrlab
This document describes JamSketch, a system that allows non-musicians to enjoy musical improvisation. It works by having users draw a melodic outline, then generates a melody in real-time to match the outline using a genetic algorithm. An experiment found that while JamSketch was inferior to experienced musicians, it performed as well or better than inexperienced musicians. Future work includes improving rhythm variation and considering relationships between pitch and rhythm.
This document presents a method for generating walking bass lines using hidden Markov models. It proposes three methods for defining hidden states and evaluates them objectively and subjectively. Method 3, which considers pitch class and metrical position, performed best by generating bass lines that were harmonically congruent, sequentially smooth, and preferred by an expert bassist over the ground truth. While effective, the approach has limitations like only handling short, simple chord progressions and more work is needed to address longer, more complex progressions and additional musical factors.
JamSketch: Improvisation Support System with GA-based Melody Creation from Us...kthrlab
This document describes JamSketch, a system that allows non-musicians to enjoy musical improvisation. It works by having users draw a melodic outline, then generates a melody in real-time to match the outline using a genetic algorithm. An experiment found that while JamSketch was inferior to experienced musicians, it performed as well or better than inexperienced musicians. Future work includes improving rhythm variation and considering relationships between pitch and rhythm.
1. The study analyzed bass melodies from Red Hot Chili Peppers albums before and after 1999 to investigate changes in the playing style of bassist Flea over time. Features related to pitch, duration, and note counts were extracted from MIDI data and used to accurately classify melodies as pre- or post-1999.
2. A second study investigated the relationship between verbal impressions of equalized vocal tones (e.g. warm, bright) and parametric equalizer settings. Participants evaluated modified vocal recordings and their responses were used to map impressions to frequency boosts and cuts.
3. The results of both studies could help develop music analysis and generation systems that incorporate stylistic changes over time or allow intuitive equalizer
The document discusses an improvisation support system called JamSketch that allows musical novices to enjoy creating music. JamSketch uses a simple drawing interface where users draw melodic outlines, and a genetic algorithm then generates a complete melody in real-time based on the outline. An experiment found the system-generated melodies were comparable to inexperienced human players but lacked human-likeness. Ongoing work aims to improve rhythm variety and allow input on smartphones or with eye tracking. The author believes systems should act as ghostwriters to support but not replace human musical ideas.
A Machine Learning Approach to Support Music Creation by Musically Untrained ...kthrlab
The document proposes a machine learning approach to support music creation by non-musicians. It describes an interface where users can edit melodies by redrawing melodic outlines or curves representing the pitch trajectory. A hidden Markov model is used to generate melodies that closely match the user's edited melodic outline while maintaining musical appropriateness, by transforming between the melody and its outline. The system aims to allow novices to input abstract musical ideas and generate pieces in an intuitive way powered by machine learning.
Extracting Melodic Contour Using Wavelet-based Multi-resolution Analysiskthrlab
The document proposes a wavelet-based multi-resolution analysis approach to extract melodic contour in a non-notewise and hierarchical manner. This approach represents melodies at different levels of resolution and abstraction through decomposition and reconstruction using the discrete wavelet transform. The approach is applied to tasks like repetition detection in melodies and measuring cognitive melodic similarity. The goals are to establish a theory of non-experts' melody cognition and develop a melody representation that is non-notewise and hierarchical.
Music Synchronizer with Runner's Pace for Supporting Steady Pace Joggingkthrlab
1) The document describes a music synchronization system that aims to help joggers maintain a steady pace by automatically adjusting the speed of the music playback based on the jogger's real-time pace.
2) An experiment was conducted with 10 participants where they jogged for 2 minutes with the system and without it in alternating trials.
3) The results showed that with the system, participants' average pace was steadier across trials and the standard deviation of their temporal pace variation was lower, indicating the system helped them better maintain a consistent speed.
Introduction of my research histroy: From instrument recognition to support o...kthrlab
This document introduces Tetsuro Kitahara and summarizes his research history in music information retrieval and automatic music generation. It describes his early work on instrument recognition in polyphonic music using probabilistic models. It then outlines his later research developing probabilistic models for computer-assisted music creation tools that allow users to generate and edit melodies and harmonies through intuitive interfaces. The document emphasizes that his recent works aim to automatically generate music from user inputs while facilitating human-computer interaction through abstract representations that hide implementation details.
1. The study analyzed bass melodies from Red Hot Chili Peppers albums before and after 1999 to investigate changes in the playing style of bassist Flea over time. Features related to pitch, duration, and note counts were extracted from MIDI data and used to accurately classify melodies as pre- or post-1999.
2. A second study investigated the relationship between verbal impressions of equalized vocal tones (e.g. warm, bright) and parametric equalizer settings. Participants evaluated modified vocal recordings and their responses were used to map impressions to frequency boosts and cuts.
3. The results of both studies could help develop music analysis and generation systems that incorporate stylistic changes over time or allow intuitive equalizer
The document discusses an improvisation support system called JamSketch that allows musical novices to enjoy creating music. JamSketch uses a simple drawing interface where users draw melodic outlines, and a genetic algorithm then generates a complete melody in real-time based on the outline. An experiment found the system-generated melodies were comparable to inexperienced human players but lacked human-likeness. Ongoing work aims to improve rhythm variety and allow input on smartphones or with eye tracking. The author believes systems should act as ghostwriters to support but not replace human musical ideas.
A Machine Learning Approach to Support Music Creation by Musically Untrained ...kthrlab
The document proposes a machine learning approach to support music creation by non-musicians. It describes an interface where users can edit melodies by redrawing melodic outlines or curves representing the pitch trajectory. A hidden Markov model is used to generate melodies that closely match the user's edited melodic outline while maintaining musical appropriateness, by transforming between the melody and its outline. The system aims to allow novices to input abstract musical ideas and generate pieces in an intuitive way powered by machine learning.
Extracting Melodic Contour Using Wavelet-based Multi-resolution Analysiskthrlab
The document proposes a wavelet-based multi-resolution analysis approach to extract melodic contour in a non-notewise and hierarchical manner. This approach represents melodies at different levels of resolution and abstraction through decomposition and reconstruction using the discrete wavelet transform. The approach is applied to tasks like repetition detection in melodies and measuring cognitive melodic similarity. The goals are to establish a theory of non-experts' melody cognition and develop a melody representation that is non-notewise and hierarchical.
Music Synchronizer with Runner's Pace for Supporting Steady Pace Joggingkthrlab
1) The document describes a music synchronization system that aims to help joggers maintain a steady pace by automatically adjusting the speed of the music playback based on the jogger's real-time pace.
2) An experiment was conducted with 10 participants where they jogged for 2 minutes with the system and without it in alternating trials.
3) The results showed that with the system, participants' average pace was steadier across trials and the standard deviation of their temporal pace variation was lower, indicating the system helped them better maintain a consistent speed.
Introduction of my research histroy: From instrument recognition to support o...kthrlab
This document introduces Tetsuro Kitahara and summarizes his research history in music information retrieval and automatic music generation. It describes his early work on instrument recognition in polyphonic music using probabilistic models. It then outlines his later research developing probabilistic models for computer-assisted music creation tools that allow users to generate and edit melodies and harmonies through intuitive interfaces. The document emphasizes that his recent works aim to automatically generate music from user inputs while facilitating human-computer interaction through abstract representations that hide implementation details.
8. Universitat Pompeu Fabra (UPF)
●
バルセロナにある公立大学
● Times Higher Education によると、 7 fastest-rising
young universities in the world の1つらしい
(Wikipediaより)
● Department de Tecnologies de la Informació i les
Comunicacions に Music Technology Group が
ある
● 3つのキャンパスに分かれており、MTGのキャンパス
は、比較的バルセロナの中心部にある
9. Music Technology Group (MTG)
音楽信号処理、音楽情報検索などの世界的拠点の1つ
Audio Signal Processing Lab. Music Information Research Lab.
Music and Multi-modal
Interaction Lab.
Music and Machine Learning Lab.
Xavier Serra Emilia Gomeź
Singing voice synthesis
Source sepration
Sergi Jordà
Reactable
Rafael Ramirez
Expressive performance
Violin
Brain-machine
music interface
PostDoc, PhD studentsも入れると50人規模のグループ
画像は各研究者・プロジェクトのWebサイトから引用
10. MTGの教育プログラム
● 3学期制
– 1学期:9月第4週~、2学期:1月第2週から、3学期:4月~
– 各学期10週間の授業期間+試験期間
– 1回の授業が2時間半のものが多い
● Master in Sound and Music Computing
– MTGが主宰するMaster course program
– 1年間full timeが基本(2年かけて修了してもいい)
– 1学期は授業中心、1学期の最後にthesis proposal発表
– 2学期からは研究中心に移行
– 学内インターンもある
23. 3) 演奏表情付け
● [Giraldo & Ramirez 2016] のモデルを使用
C
Feature extraction
for each note
k-NN, multilayer perceptron, ...
onset deviation
duration ratio
energy ratio
(Duration, onset, prev. duration,
next duration, prev. interval,
next interval, namour, chord, etc.)
31. 会議参加
● 31 August-3 September 2016
– Sound and Music Computing Conference (SMC 2016)
– Hamburg, Germany
● 6-9 September 2016
– International Conference on Artificial Neural Networks (ICANN
2016)
– Barcelona, Spain
● 19-23 September 2016
– European Conference on Machine Learning and Principles and
Practice of Knowledge Discovery (ECML-PKDD 2016)
– Incl. “International Workshop on Music and Machine Learning
(MML 2016)” (ポスター発表)
– Reva del Garda, Italy
32. ● 28 November-2 December 2016 (招待講演)
– Joing Meeting of Acoustical Society of America (ASA) and
Acoustical Society of Japan (ASJ)
– Honolulu, USA
● 4 December 2016 (ポスター発表)
– Seminar on Music Knowledge Extraction Using Machine Learning
– Barcelona, Spain
● 5-10 December 2016
– Annual Conference on Neural Information Processing Systems
(NIPS 2016)
– Incl. “Constructive Machine Learning Workshop” (ポスター発表)
– Barcelona, Spain
● 20 December 2016 (ポスター発表)
– Digital Music Research Network One-day Workshop 2016
– London, UK
33. 研究室訪問
● 7 November 2016
– IRCAM, Paris, France
– Dr. Tsubasa Tanaka
● 19 December 2016
– Queen Mary University
of London
– London, UK
– Dr. Eita Nakamura
● 28 December 2016
– Academia Sinica
– Taipei, Taiwan
– Dr. Yi-Hsuan Yang
● 27 January 2017
– University of the Basque
Country
– San Sebastian, Spain
– Prof. Darrell Conklin
● 2 February 2017
– Sony CSL Paris, Paris, France
– Dr. Francois Pachet
– Dr. Pierre Roy
● 3 February 2017
– Open University
– Milton Keynes, UK
– Prof. Simon Holland
– Prof. Robin Laney
34. SMC 2016
● Sound and Music Computing Conference
● ISMIRなどに比べるとレベルは高くない印象
35. ECML-PKDD 2016
● ECML: Europian Conf. on Machine Learning
● PKDD: Principle and Practice of Knowledge Discovery
● Google系の招待講演が目立った
36.
37. MML 2016
● Int'l Workshop on Music and Machine Learning
● ECML-PKDDのワークショップとして実施
●
ポスター発表も行った
38. Joint Meeting of ASA/ASJ
● 10年に1度の日米音響学会合同の大会
● Music Signal Processingセッションで招待講演
39. NIPS 2016
● Neural Information Processing System
●
人工知能/機械学習ブームにともなって参加者急増中
● 当日参加受付一切なし、受付時にID提示の異例の体制
43. 研究室訪問:University of the Basque Country
● 訪問先:Prof. Darrell Conklin
● Music generation by transformation
M
A A’
M’
● 題材として trans music の chord progression
● San Sebastian はとても美しい街で、食べ物が旨い
44.
45. 研究室訪問:Sony CSL Paris
● 訪問先:Dr. Francois Pachet, Dr. Pierre Roy, et al.
● Flow Machines
– 自動作曲Webアプリケーション
– 種となる楽曲(群)を指定すると、マルコフモデルを学習
– 生成旋律の種楽曲への類似度をパラメータで制御可能
– 楽音のレンダリングも実装(プラグインがサーバ上で動作)
– 歌唱音響信号を指定し、その歌声で歌わせることも可能
● DeepBach
– 四声体和声の自動生成
– LSTMで動作
46. 研究室訪問:Open University
● 訪問先:Prof. Simon Holland, Prof. Robin Laney
● Music Computing Lab.にて様々なプロジェクト実施
– Haptic Bracelets
– Harmony Space
– Modeling of polyrhythm perception
– Support of holding the violin for beginners
– Music programming language for beginners
– Embodied cognition theory improves musical
instruments
– Generating music for games
67. 私が覚えたスペイン語のすべて
● ¡Hola! (Hello!)
● ¡Gracias! (Thanks!)
● ¡Adios! (Bye!)
● ¡Muy bien! (Very good!)
● ¡Perfecto! (Perfect!)
あいさつ篇 質問・お願い篇
● ~, por favor.
(~, please.)
● ¿Hay ~?
(Is there ~?)
● ¿Donde esta ~?
(Where is ~?)
飲食店で便利な単語篇
● esto (this)
● aquí (here)
● para llevar (take away)
● uno mas (one more)
68. ● patata (potato)
● hamburguesería
(burger)
● pan (bread)
● atún (tuna)
● pollo (chicken)
食べ物篇
飲み物篇
● cerveza (beer)
● cerveza sin alcohol
(beer without alcohol)
● vino tinto (red wine)
● vino blanco (white wine)
● cava (sparkling wine)
● café (coffee)
● americano (American)
● café con leche
(coffee with milk)
● agua sin gas
(water without gas)
● agua con gas
(water with gas)