The Passport to Paradise gallery highlights the bold, visual images found all over Dakar by focusing upon the urban visual culture of the Mourides, a Senegalese Sufi movement centered upon the life and teachings of a local saint named Sheikh Amadou Bamba.

Read our recently published guides to best practices in digitizing text and multimedia resources. Download them today from our "Best Practices" site.

Digitizing Speech Recordings for Archival Purposes

Working Paper on Digitizing Audio
for the Nation Gallery of the Spoken Word and
the African Online Digital Library

MATRIX: The Center for Humane Arts, Letters and Social Sciences Online

at Michigan State University

Goals

The bulk of the "National Gallery of the Spoken Word" (NGSW) [1] digitization process involves digital processing of speech signals. All of the sound archives in question exist in the analog format. They are usually stored on a variety media, including wax cylinders, LPs, reel-to-reel tapes, and audio cassettes. The content of those archives is, by and large, speech. The primary purpose of digitization is preservation. However, the digital format of audio recordings lends itself to a variety processes ranging from delivery over the Internet to speech recognition-based searching. The digitization process, obviously, is not a simple one and it poses a number of challenges that the archivist needs to address. Some of those challenges are due to the variety of original media, media quality, and recording quality. Other concerns include the choice of the digital file format, its characteristics, the storage medium, as well as the choice of appropriate hardware, software, and digitization procedures. Finally, due to the large volume of the collections to be digitized, work flow and quality control issues have to be carefully addressed. This paper outlines some of the key concepts in speech digitization and a working model for digitizing speech for archival purposes.

Choosing appropriate digitization standards

For the purposes of the analog-to-digital conversion of most of VVL audio recordings, we have chosen to use a 44,100 Hz sampling rate and a 16-bit resolution. The details of this decision, as well as the process itself, are delineated below.
The choice of appropriate digitization standards seems to be influenced by two distinct factors - the technological and the archival. Each of these factors poses a different set of questions and challenges to the project. When deciding on particular specifications of sampling rate, quantization, hardware, etc. our primary technological goal is to provide a digital copy that closely matches the analog original. In the technical sense, we need to establish a process that, minimally, reconstructs the entire frequency response of the original while adding as little of the so-called "digital noise" as possible. To achieve this goal, it seems to be sufficient to use the 44,100 Hz sampling rate and a 16-bit resolution. The former ascertains that we capture the entire audible frequency range of the original (see Nyquist theorem below), while the latter, gives us a fairly good, 96 db SNR (signal to noise ratio). From the archival standpoint, it is our desire to preserve as much information (both speech and non-speech) of the original as possible. The technical specifications mentioned above promise to fulfill this goal.

In the ideal world, the discussion on choosing appropriate standards could end right here. However, there are a few other issues involved. Current technology makes it possible to use higher sampling rates and resolution. One could fairly easily sample at 96,000 Hz and a 24-bit resolution. This would result in a much increased frequency response of the digital waveform ď from 0 to 48,000 Hz, and a dramatically improved SNR of 144 dB. At first glance, this technology appears to provide a very simple solution. After all, it is our goal to preserve as much information as possible, and using such high specifications does exactly that. The digital audio file captured at 96,000 contains over twice as much information as the one sampled at 44,000 Hz. The question, however, is whether the extra information included in the 96,000 Hz file is worth preserving.

NGSW recordings can be easily characterized as limited bandwidth. First of all, they are all analog speech recordings. Speech, as such, is a rather narrow-bandwidth medium. In fact, the highest frequency speech sounds, voiceless fricatives, such as /th/ and /f/ have an upper frequency of about 5,000 ď 6,000 Hz, depending on the age and gender of the speaker. Thus, any information in the channel that is above the 6,000 Hz is, without a doubt, non-speech information. Of course, from the archival standpoint, we may want to preserve this information, too. That could be any "environmental", ambient sound present in the recording. Second, most NGSW recordings have been mastered on a limited-bandwidth analog equipment that was not capable of capturing any of the high frequency information in the first place. The Nyquist frequency in such recordings varies, but it hardly ever goes beyond the audible range of 20-20,000 Hz.

In light of these facts, the higher specifications help us capture nothing but silence. The spectrogram below illustrates a phrase: /yacIyesteshmI/ converted to digital at 96,000 Hz. As we can see, acoustic content of this file occupies only the lower 1/3 of the spectrum. Anything above that is silence.

The question of whether we should adopt the higher specifications, however, still remains to be answered. Beside that facts mentioned above, it has been suggested in psycho-acoustic literature that higher sampling rates do have a clear perceptual effect. Most of the experiments done so far have focused on music, especially recordings originally captured at 96,000 Hz. It still remains to be seen whether this effect is present in speech recordings, especially those converted from analog media. One last remaining issue is the effect of anti-aliasing filters used in a process clocked at 96,000 Hz. Arguably, one could apply a filter with a slope much less steep than in the process clocked at 44,100 Hz. NGSW sound engineers are currently in the process of researching the perceptual effects of higher sample rates and more gradual anti-aliasing filters. The decision of whether to use 96,000.24-bit specifications largely depends on the results of those tests. In making the decisions, NGSW will also consider the issues of data storage and migration, as well as available processing power. It must be noted that higher digitization standards require considerably more storage space and processing power.

Basic Concepts

Digitization

Digitization is a process of converting an analog, or continuous, waveform to digital form by an analog to digital converter (ADC). The sound pressure waveform is a continuously varying signal. It changes from instant to instant, and as it changes between two values, it goes through all values in between. Using an ADC, the signal is sampled at a set of equally spaced times. The sample value is a number equal to the signal amplitude at the sampling instant. Computers always represent numbers internally in the binary number system, where each binary position can be one of the two binary digits, 0 or 1. Each binary digit is called a bit, and a collection of eight bits is called a byte.

The astonishing thing about the ADC process is that, if done properly, every detail of the original signal can be captured. The original continuous waveform can be reconstructed exactly and, more importantly, powerful digital signal processing can be applied to the digital representation of the signal. Such a representation can then serve a variety of purposes, such as streaming, watermarking, speech recognition, etc. A vast number of effects, such as noise reduction, bass emphasis, normalization, compression, etc. can be achieved by applying a mathematical transformation called "a digital filter." After processing , the resulting continuous signal can be reconstructed without any degradation. The process of converting from digital back to a continuous signal is called digital to analog conversion, or DAC.

When any signal is transmitted over a communications medium, it is inevitably distorted to some degree. The received signal, therefore, differs to some extent from the originally transmitted signal. For an analog signal, even small amounts of distortion lead to noise that is often impossible to eliminate. Examples of noise include "hiss", "static", and the poor quality (limited bandwidth) of signals transmitted over telephone circuits. For digital signals, the same signal distortion factors are present. However, since the signal at any instant of the represents a number, rather than an analog signal level, it is necessary only to unambiguously recognize the number that was transmitted. This choice is substantially eased by the fact that digital signals are binary. As long as noise is small enough so that the choice between these two outcomes is not confused, the received signal will represent the same numerical value as the transmitted signal. In fact, additional coding techniques, such as parity check, are used to detect and correct errors that might occur in the process. The result is unparalleled quality and robustness of digital signals.

The sampling theorem

It is not obvious that an exact reconstruction of an analog signal should be possible, since a complete continuous signal is replaced by a finite number of samples taken at equal time intervals. The problem is to have complete information between the samples. The answer lies in the mathematical result called the Sampling Theorem. In short, the sampling theorem states that if a band-limited signal is sampled at a rate greater than twice the highest frequency in the signal (the so-called Nyquist frequency), no information is lost and the original signal can be exactly reconstructed from the samples. Acoustic signals that humans can hear lie in a limited range of about 20 to 20,000 Hz. Thus, intuitively, in order to exactly reconstruct the original analog signal one should use the sample rate of at least 40,000 Hz. This is, indeed, true. It is recommended to use the standard, CD quality, sample rate of 44.1 KHz for all of spoken word digitization projects. It is important to note that sampling at a rate lower than twice the Nyquist frequency results in a phenomenon called aliasing, whereby the resulting digital signal lacks a significant number of bits of information and it may be severely distorted by the aliased components, making it audible as noise. Finally, even though the speech signal usually does not contain any information above 7 kHz, and, theoretically, the sample rate of 16 KHz should be sufficient to capture all details of the signal, it is nevertheless recommended to use the sample rate of 44.1 kHz for all of AD conversion for archival purposes.

Quantization

For the sampling theorem to apply exactly, each sampled amplitude value must exactly equal the true signal amplitude at the sampling instant. Real ADCs do not achieve this level of perfection. Normally, a fixed number of bits (binary digits) is used to represent a sample value. Therefore, the infinite set of values possible in the analog signal is not available for the samples. In fact, if there are R bits in each sample, exactly 2R sample values are possible. For high-fidelity applications, such as archival copies of analog recordings, 16 bits per sample (65, 536 levels), or a so-called 16 bit resolution, should be used. The difference between the analog signal and the closest sample value is known as quantization error. Since it can be regarded as noise added to an otherwise perfect sample value, it is also often called quantization noise. The effect of quantization noise is to limit the precision with which a real sampled signal can represent the original analog signal. This inherent limitation of the ADC process is often expressed as a Signal-to-Noise ratio (SNR), the ratio of the average power in the analog signal to the average power in the quantization noise. In terms of the dB scale, the quantization SNR for uniformly spaced sample levels increases by about six dB for each bit used in the sample. For ADCs using R bits per sample and uniformly spaced quantization levels, SNR = 6R - 5 (approximately). Thus, for 16-bit encoding about 91 dB is possible. It is 20 to 30 dB better than the 60 dB to 70 dB that can be achieved in analog audio cassette players using special noise reduction techniques.

The WAV file format

The WAV file format is recommended for storing digital versions of speech recordings. WAV files are uncompressed, thus preserving all bits of information recorder in the AD process. It is also widely used and easy to process and convert to a variety of streaming formats.
The WAV file format is a variant of the RIFF format for data interchange between programs. This format was designed so that data in a file is broken up into self-described, independent "chunks". Each chunk has a prefix which describes the data in that chunk. The prefix is a four-character chunk ID which defines the type of data in the chunk, followed by a 4-byte integer which is the size of the rest of the chunk in bytes. The size does not include the 8 bytes in the prefix. The chunks can be nested. In fact, a RIFF file contains a single chunk of type "RIFF", with other chunks nested inside it. Therefore, the first four bytes of a WAV file are "RIFF", and the four bytes after that contain the size of the whole file minus 8 bytes. After the RIFF header is the WAV data, consisting of the string "WAVE" and two important chunks: the format header and the audio data itself. There may also be other chunks in a WAV file that contain text comments, copyrights, etc., but they are not needed to play the recorded sound. The header and the audio data can be easily separated to facilitate migration to any other format, if necessary.

Hardware

A/D converter

We recommend using a professional-level, hardware platform, preferably integrated with the software, that includes reliable signal acquisition, analysis, and playback, such as Kay Elemetrics' Computerized Speech Lab (model 4300B). Such hardware should include 16-bit input/output, auto calibration, sophisticated signal conditioning and anti-aliasing filters to take full advantage of the 16-bit signal representation. Additionally, the hardware should be able to offer a wide assortment of frequency ranges, XLR (balanced) inputs, DAT pass-through, and powerful digital signal processing circuitry for fast signal analysis. Unlike systems built around generic plug-in multimedia sound cards (which were designed primarily for game sound output, not input), such hardware usually offers input signal-to-noise performance typically 20-30dB superior to generic plug-in sound cards. There are a number of superior quality A/D converters available on the market. It is important, however, that the A/D converter meet the above specifications.
Hardware-based ADC usually offers the following features:
  • Superior S/N specifications due to isolation of analog input/output from computer
  • Calibrated input and sensitivity for measurement of absolute level (SPL)
  • High gain preamplifier to accommodate low-level input signal levels
  • 2- or 4-channel input Digital (e.g., DAT) input/output
  • XLR (balanced) microphone input for protection from noise
  • Online feedback and adjustment of input level
  • User-selected AC/DC coupling for all channels
  • Wide selection of frequency ranges
  • Manual and computer control of input and output level
  • Auto-calibration of hardware Real-time indicator of signal overloading
The external module also performs the initial signal conditioning and A/D conversion. The external module isolates the low-level (e.g., microphone) analog input signals from common sources of noise in PC-based sampled data systems: the switching noise of digital signals, power grounding, and unshielded power supplies. "It seems like there is so much noise and interference going on inside a computer that it■s necessary either to use a digital I/O or have some sort of external breakout box with A/D converters instead of the «traditional sound■ card." (Ken Lee, New Media, April 14, 1997, pp. 33-44.)

PCI Sound Card

The analog signal is converted to digital by the external module. Then it is captured by an internal PCI sound board. It is important to avoid using standard multimedia sound cards. It is recommended to use a professional-level sound card that meets the following specifications:
  • PCI Interface
  • 8 to 24 bit resolution
  • variable sample rates, including 11.025kHz and 44.1kHz
  • Analog two channel in/out via 1/4" TRS connectors
  • S/PDIF digital in/out via gold-tipped RCA connectors
  • +4/-10dB balanced/unbalanced operation
  • 4 channel operation using both analog and digital ins and outs simultaneously
  • Analog Input Dithering

Computer

  • A top-of-the-line PC or Mac that meets the following specifications:
  • 500MHz or higher processor
  • 256 MB of RAM
  • High-capacity, fast HDD (preferably 2 Ultra 160 SCSII drives with a RAID controller card)
  • Fast CD Writer
  • Fast DVD Drive
  • IEEE 1394 card for digital in/out (optional)

Software

  • A sophisticated sound editor is recommended. It should meet the following specifications:
  • Multiple Document Interface for editing multiple files in one session
  • Unlimited file size editing
  • Real-time graphs (amplitude, spectrum, bar, and spectrogram) (optional)
  • Real-time fast forward and rewind playback
  • Numerous effects (distortion, Doppler, echo, filter, mechanize, offset, pan, volume shaping, invert, resample, transpose, noise reduction, time warp, pitch, and more)
  • Supports many file formats (WAV, AU, IFF, VOC, SND, MAT, AIFF, MP3, OGG, and raw data) and can convert to/from these formats
  • Drag-and-drop cue points
  • Direct sample editing with the mouse

Digitization process overview

The digitization process, is a rather straightforward one. With a little bit of training and correct preset parameters, it could be done by anyone with basic computing skills.

Summary of A/D conversion settings:

  1. Sample rate: 44.1 kHz
  2. Resolution: 16 bit
  3. A/D converter settings: use plain signal with no correction. Use normalization if necessary.
  4. Input level: adjust and monitor levels to avoid signal overload. Digitization is very prone to clipping on overload. The process of setting levels is different for digital and analog.
  5. Save as: *.wav


[1] "The National Gallery of the Spoken Word" is a National Digital Libraries Initiative, Phase II project to create a significant, fully searchable, online database of spoken word collections spanning the 20th century. The first large-scale repository of its kind, the NGSW will provide storage for these digital holdings and public "exhibit" space for the most evocative collections.

Sampling from L'Institut Fondemental d'Afrique Noire (IFAN)

Phil Curtin Collection

Collection Boubacar Barry

Collection Charles Becker: Recherches et documents sur le Sida

Photographs from “Passport to Paradise’: Sufi Arts of Senegal and Beyond

Mosques of Bondoukou

Futa Toro, Senegal and Mauritania

Build Your Own Gallery

Galleries | Research | Institutions | Best Practices | Search | Home