1
Prototype of a Radio-On-Demand Broadcast Receiver with Real-Time Musical Genre Classification B. HAPPI TIETCHE 1 , O. ROMAIN 4 , B. DENBY 2 , L. BENAROYA 2 , F. DE DIEULEVEULT 3 , B. GRANADO 4 , G. WASSI 4 , H. KHEMIRI 5 , G. CHOLLET 5 , D. PETROVSKA 5 , R. BLOUET 6 , K. HACHICHA 1 AND S. VIATEUR 1 1. Laboratoire Informatique de Paris 6 – LIP6 – Université Pierre et Marie Curie 2. Université Pierre et Marie Curie, Laboratoire SIGMA, ESPCI-ParisTech 3. Laboratoire LIST – CEA 4. Laboratoire ETIS – Université Cergy Pontoise 5. Institut Mines Télécom 6. YACAST [email protected], [email protected], [email protected] Software Defined Radio applied to broadcast media allows to imagine digitizing an entire broadcast band, decomposing it into its component channels, demodulating these to produce parallel real-time audio streams, and applying indexing or other audio processing algorithms, all in a single programmable device. Objective Radio On Demand Speech Music Genre Song ID ctc. Word spot Speaker ID ctc. Station 1 Multimedia Database / User Interface Speech Music Genre Song ID ctc. Word spot Speaker ID ctc. Station N LNA + ADC Audio Indexing Processor Multimedia Interface Channelizer / Demodulator Antenna Audio Indexing Prototype HMI Stratix Altera FPGA : EP2S180 - IP WOLA - Resampling - Cordic • Multi-core - GMM - HMI ALUTs 67999/143520 (45%) Registers 64950/143520 (36%) Total registers 64950 Memory in bits 2275985/9383040 (24%) DSP block 9 bit 464/768 (59%) Features Algorithms Bandpass sampling @ 87MHz • Resampling @ 51.2MHz • 4 WOLA Filter with 128point FFT running @ 51.2MHz • // Cordic Parallel demodulator Musical Genre Classification MFCC to extract acoustic feature • Gaussian Mixture Model Classifier • Training on GTZAN (http://marsayas.info) and French radio datase • Decision based on the maximum of the three scores • 96.3% of accuracy Classical Jazz Variety Classical 86% 4% 1% Jazz 10% 93% 0% Variety 4% 3% 99% Confusion matrix Core 2 duo Q9400@ 2.66GHz Core i7 2600@ 3.4GHz 1 vector 699ms 457ms 51 channels 35.67s 23.307s Processing time on multi- core

Prototype of a Radio-On-Demand Broadcast Receiver with

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Prototype of a Radio-On-Demand Broadcast Receiver with

Prototype of a Radio-On-Demand Broadcast Receiver with Real-Time Musical Genre Classification

B. HAPPI TIETCHE1, O. ROMAIN4, B. DENBY2, L. BENAROYA2, F. DE DIEULEVEULT3, B. GRANADO4, G. WASSI4, H. KHEMIRI5, G. CHOLLET5, D.

PETROVSKA5, R. BLOUET6, K. HACHICHA1 AND S. VIATEUR1

1. Laboratoire Informatique de Paris 6 – LIP6 – Université Pierre et Marie Curie 2. Université Pierre et Marie Curie, Laboratoire SIGMA, ESPCI-ParisTech

3. Laboratoire LIST – CEA 4. Laboratoire ETIS – Université Cergy Pontoise

5. Institut Mines Télécom 6. YACAST

[email protected], [email protected], [email protected]

Software Defined Radio applied to broadcast media allows to imagine digitizing an entire broadcast band, decomposing it into its component channels, demodulating these to produce parallel real-time audio streams, and applying indexing or other audio processing algorithms, all in a single programmable device.

Objective Radio On Demand

Speech Music

Genre Song ID

ctc.

Word spot Speaker ID

ctc.

Station 1

Multimedia Database / User Interface

Speech Music

Genre Song ID

ctc.

Word spot Speaker ID

ctc.

Station N LNA

+

ADC

Audio

Indexing

Processor

Multimedia

Interface

Channelizer / Demodulator

Antenna Audio Indexing

Prototype

!!

HMI

•  Stratix Altera FPGA : EP2S180 - IP WOLA

- Resampling

- Cordic •  Multi-core

- GMM

- HMI

ALUTs 67999/143520 (45%) Registers 64950/143520 (36%) Total registers 64950 Memory in bits 2275985/9383040 (24%) DSP block 9 bit 464/768 (59%)

Features

Algorithms

•  Bandpass sampling @ 87MHz

•  Resampling @ 51.2MHz

•  4 WOLA Filter with 128point FFT running @ 51.2MHz

•  // Cordic

Parallel demodulator Musical Genre Classification •  MFCC to extract acoustic feature

•  Gaussian Mixture Model Classifier

•  Training on GTZAN (http://marsayas.info) and French radio datase

•  Decision based on the maximum of the three scores

•  96.3% of accuracy

Classical Jazz Variety Classical 86% 4% 1% Jazz 10% 93% 0% Variety 4% 3% 99%

Confusion matrix

Core 2 duo Q9400@ 2.66GHz Core i7 2600@ 3.4GHz 1 vector 699ms 457ms 51 channels 35.67s 23.307s

Processing time on multi-core

! ! !