Skip to the content ForoHackers
Log inJoin

Speech: start here

Speech
#2847 · · Quote · Report

Speech technology covers turning spoken audio into text, text into audio, and telling who is speaking or in what language. It powers captions, dictation, voice interfaces and accessibility tools. Speech is a signal that varies with speaker, accent, noise and microphone, so the work is half signal processing and half learning.

Key concepts

  • Audio basics: sampling rate, bit depth, framing and the Fourier transform, which turns a waveform into a spectrum.
  • Features: spectrograms and mel-scale representations, the usual input to models.
  • Speech recognition: mapping audio to text, with an acoustic model plus a language model, or one end-to-end network.
  • Speech synthesis: text to waveform, via an acoustic model that predicts a spectrogram and a vocoder that renders it.
  • Speaker tasks: identification, verification and diarisation, which marks who spoke when.

Vocabulary

  • Phoneme: the smallest sound unit that distinguishes words in a language.
  • Spectrogram: a picture of frequency content over time.
  • Word error rate: substitutions, deletions and insertions divided by the number of reference words.
  • Forced alignment: matching a known transcript to the exact moments in the audio.

What to study, in order

  1. Record a sound, plot its waveform and spectrogram, and see how vowels differ.
  2. Study sampling, aliasing and filters in a signal-processing course or text.
  3. Build a keyword detector on mel features with a small network.
  4. Run a recognition system on recordings of different voices and compute the word error rate.
  5. Read how synthesis works, then test how an output fares with names, numbers and noisy input.

Related tags

#signal-processing #nlp #deep-learning #accessibility #multimedia

You are not logged in, so this goes up anonymously: no account, no name and no address is stored with it, and it may carry no link and no image. Log in or join to post under your name. Our web server's logs keep the address and time of every request for up to 14 days; to keep yours out of them, use Tor. To reach the staff in confidence, use the drop. Sign as a returning source (a key kept only in this browser)

Sending takes a few seconds here: your browser first works out a proof of work against spam.

Reply