AI Glossary / Natural language and speech / Intermediate
A-D conversion
Definition
A-D conversion means analog-to-digital conversion: turning a continuously varying physical signal into a sequence of digital values. For speech, a microphone first produces an electrical representation of sound, and a converter samples and quantizes that signal. The result is a numeric waveform that speech-recognition and other audio-processing systems can store, transmit, and analyze.
What it means
An analog waveform varies continuously over time and amplitude. A digital representation makes two choices: when to measure it and which numeric values are available to describe each measurement. Sampling discretizes time; quantization discretizes amplitude.
These choices are different from speech recognition. Conversion produces samples, not words or meanings. They are also different from an audio codec's later compression decisions. A high-quality digital recording may still contain background noise, an unsuitable microphone position, or speech the recognition model handles poorly.
How it works
The input signal is conditioned to suit the converter's range, and an anti-aliasing filter limits frequencies before sampling. The sampling rate specifies how many amplitude measurements are taken per second. Quantization maps each measured amplitude to one of a finite collection of representable levels, which can then be encoded as bits.
Sampling theory relates the rate to the signal's bandwidth. Ideal reconstruction assumptions require an appropriately band-limited signal and a sampling rate exceeding twice its highest retained frequency. Real systems also need filter transition margins and account for noise and hardware limitations; “two samples per cycle” is not a universal promise of accurate recording.
Bit depth controls the number of available amplitude codes. Gain must avoid clipping, where the signal exceeds the converter's range. Increasing stored precision later cannot recover detail already lost through clipping, aliasing, or coarse quantization during capture.
Three practical examples
1. Voice interfaces: digitize microphone signals before a speech model extracts features or directly processes the waveform.
2. Meeting transcription: capture one or more audio channels with known sample rates so that timing and speaker-channel information remain interpretable.
3. Equipment monitoring: convert an analog vibration or sound signal into samples for a model that detects patterns in machine operation.
Think of it this way
Think of measuring a moving water level at regular moments and recording each result using the markings on a ruler. The measurement schedule resembles sampling, and the ruler's available increments resemble quantization. The analogy has limits: audio signals can change far faster, and frequencies above the sampling system's limit can appear falsely as lower frequencies rather than merely being omitted.
Worked example
Suppose a recording uses 16,000 samples per second, 16 bits per sample, and one channel. Ten seconds contains 160,000 samples. Uncompressed sample data therefore occupies 160,000 × 16 = 2,560,000 bits, or 320,000 bytes. File headers and other metadata are additional.
A 20-millisecond interval contains 16,000 × 0.020 = 320 samples. Two channels at the same settings double the sample-data size, but do not double the recording's duration.
For a separate simplified quantization example, imagine a uniform 12-bit converter spanning a two-volt input range. It has 2^12 = 4,096 codes, with nominal interval width 2/4,096 volts, approximately 0.488 millivolts. Under an ideal rounding model and away from overload, the quantization error is bounded by roughly half an interval, about 0.244 millivolts. Actual hardware noise and code conventions are additional considerations.
Finally, sampling an unfiltered 5-kilohertz sinusoid at 8 kilohertz can make it appear as a 3-kilohertz component. This alias follows from the sampled sinusoid's frequency ambiguity. A later speech model cannot reliably infer the original frequency from those samples alone. The remedy is appropriate filtering and capture design, not merely changing the file's declared sample rate. These are hand-checked numerical illustrations.
Illustrative case study
Fictional case: Pine Transcripts receives audio from two recording devices. One pipeline labels every file as 16 kilohertz, although one device actually captured 8-kilohertz samples. Playback speed and timing become wrong because the samples are interpreted using incorrect metadata.
The team preserves the original capture rate, performs an explicit resampling operation when needed, and verifies duration on a known recording. It also checks for clipping rather than assuming a larger bit-depth export repairs damaged audio. The scenario describes an input-quality investigation, not a measured transcription improvement.
Common misconceptions and mistakes
Sample rate and bit depth describe different dimensions. A higher sample rate does not directly increase the number of amplitude levels, while more bits do not extend the captured frequency band.
Upsampling creates a representation at a higher rate through interpolation; it does not restore frequencies removed by the original recording system. Saving clipped audio with more bits likewise does not reconstruct the flattened peaks.
A digital microphone still involves analog-to-digital conversion internally. Conversely, a software function reading an already digital audio file is decoding or processing existing samples, not performing a fresh physical A-D conversion.
Business use
For a speech project, document capture rate, sample encoding, channel layout, gain behavior, and any resampling. Check that the model's input assumptions match the actual data rather than trusting filenames or default settings.
Evaluate the whole capture path under representative conditions. Storage calculations help planning, but intelligibility and recognition quality also depend on the acoustic environment, microphone characteristics, processing, and model suitability. More data per second is not by itself a guarantee of better business results.
People also ask
Does A-D conversion turn speech into text?
No. It turns an analog signal into digital samples; recognition is a later task.
Is quantization always audio compression?
No. Quantization is part of digital representation, while codecs may add further compression operations.
Can changing sample-rate metadata repair an incorrectly captured signal?
No. Metadata must describe the samples correctly; resampling is a separate operation and cannot recover lost information.
Check your understanding
How many samples are in one second of mono audio captured at 16 kilohertz?
Answer: 16,000. Bit depth determines how many bits encode each sample, not the number of samples per second.
Related terms
- Sampling
- Nyquist frequency
- Aliasing (factorial experiments)
- quantization
- sample rate
- bit depth
- aliasing
- pulse-code modulation
- speech waveform
Learn next
compare sampling and quantization errors, then trace how a digital speech waveform becomes short-time acoustic features.
More in Natural language and speech
- AAC - AAC stands for augmentative and alternative communication: ways of communicating that supplement or replace speech when needed.
- Absolute position - Absolute position identifies where a token occurs within a sequence using its own position index, such as first, second, or third.
- Abstract word - An abstract word expresses a meaning that is less directly tied to a specific sensory or physical experience, such as an idea, relation, quality, or mental...
- Accented syllables - Accented syllables are syllables that bear a pitch accent in an utterance and are perceived as prosodically prominent.
- Acoustic model - Definition: An acoustic model is a component of a speech-recognition system that relates recorded speech signals to linguistic units or their internal states.
- Activity expressions - Activity expressions describe situations that unfold over time without building a particular completion point into their meaning.
Sources
Primary source: Jurafsky and Martin, Speech and Language Processing, Chapter 15, sampling and quantization: https://stanford.edu/~jurafsky/slp3/15.pdf . Relevant passages were read. The original full-book URL was not retrievable; this chapter supplies the same topic. The explanation qualifies simplified sampling claims with bandwidth and filtering conditions. Examples are original.
Previous: A/B testing Next: A-law algorithm