Explorez tous les épisodes du podcast Between the Signs: The Speech Communication Podcast

The final episode. Speech is fast, expressive and available almost anywhere, and it leaves hands and eyes free — but it is serial and fleeting, and no improvement in speech technology will make a spoken list of twelve train connections memorable. So systems learn to combine channels: what a camera can add that a microphone cannot, why gaze reveals attention but rarely carries a command, and why merging channels is mostly a question of timing. The series closes where it began, with the observation that human communication was never purely acoustic, and that technical systems are still catching up.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

Teaching a machine to speak means inventing everything the text leaves out — and getting it right, because listeners use melody and rhythm to repair what they did not quite catch. This episode traces four generations of talking machines, from modelling the human speech apparatus to gluing together recorded fragments to learning the whole mapping statistically. Then the harder problem: a conversation is not a sequence of sentences. What a dialog system has to keep track of, why its most important skill is talking about the conversation itself, and why confirming what it understood is a choice between being safe and being bearable.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

The first of two episodes on the largest chapter of the lecture. Why recognizing speech cannot possibly work by looking things up: there are no gaps between spoken words, every sound is smeared into its neighbours, and the same sentence is never twice the same signal. The way out was to stop matching and start betting — to ask which sequence of words would most probably have produced what was just heard. From there: models that could have said it, features built deliberately to imitate the ear, language models as a machine’s version of world knowledge, and the more recent shift to networks that learn from unlabelled audio. Plus an honest look at where recognition still struggles.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

How a phone call survives being squeezed, told as three ideas rather than a catalogue of codecs. Spend bits where they do the most good, which is why quantization steps are fine near zero and coarse further out. Do not send what the receiver could have guessed, and shape the unavoidable error so that it hides underneath the speech itself — masking, used as a design tool. And, at the extreme, stop sending the signal at all and send the recipe instead, at the price of a voice that is intelligible but no longer anyone in particular. The episode closes with the compromise that made mobile telephony possible, and with a warning about trusting the obvious quality metric.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

Hearing is not measurement. This episode follows sound through an ear that filters by direction, solves the problem of getting sound from air into liquid, and performs a frequency analysis mechanically, by letting a travelling wave peak at a different place for every frequency. Then the part that matters most for engineers: two tones of the same physical level are not equally loud, a quiet sound next to a loud one is simply not there at all, and the ear pools what it hears into a couple of dozen bands. Which leads to the sentence the next chapter depends on: what cannot be heard does not have to be transmitted.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

A detective problem, and three ways to solve it. How a spectrum is measured in the first place, and why it helps to build the analysis filters so that they resemble the ear. Why cutting a short slice out of a signal changes what you see. Then the trick worth remembering: the logarithm turns a product into a sum, which is how the wiggles from the vocal folds can be separated from the shape imposed by the mouth — the cepstrum, complete with the delightful vocabulary the field invented for it. And finally the idea of predicting a speech sample from the ones just before it, so that what remains is the excitation a microphone can never reach.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

Where the properties of a speech signal come from. Two ways of making a sound, one of them driven by a lovely piece of physics at the vocal folds, and the observation that whispering replaces the voiced source entirely without costing a single word — so the words apparently do not live in the source. Then the filter, where they do live: a tube of some seventeen centimetres whose plain acoustics predict the resonances we actually measure. The episode ends with the idea the rest of the lecture is built on, the source-filter model, and with the pleasing fact that the physical model of the vocal tract and the signal-processing model turn out to be the same thing.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

There is no single right way to describe a speech signal, and each way reveals something an engineer can use. As a shape over time, as a mixture of frequencies, as a random process, and as a stream of numbers. This episode is about why engineers reach for frequencies, why speech has to be cut into slices of a few tens of milliseconds before anything can be measured, and why the resulting picture can be sharp in time or sharp in frequency but never in both. Along the way: what actually sits in a slice of speech, and why the fact that speech is partly predictable is the seed of everything that follows in the coding chapter.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

The opening episode sets up the whole lecture. Speech as a system of signs, and the uncomfortable insight that a system can transmit information perfectly and still fail to transmit meaning. The building blocks of language, from the smallest units that change meaning to the melody and rhythm that tell a listener what was a question and what mattered. Two models of communication, one of them missing the human entirely. And a first look at everything that can go wrong along the way: noise, echo, delay, lost packets, or simply no shared code between the two sides.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.

A short introduction to the series. Each episode gives the big picture of one chapter of the lecture and answers a single question above all: why does this topic matter for speech communication? The details stay where they belong, in the lecture and the script. This first episode explains how the episodes are meant to be used, previews the journey ahead, and is open about how the series was produced.

This episode was generated with artificial intelligence (NotebookLM) on the basis of the lecture script, and the voices are synthetic. In case of doubt, the lecture and the script are authoritative.