Vocoder: how a voice is rebuilt from a description
Jaseem Abbas
- Published
What is a vocoder?
A vocoder rebuilds the sound of speech from a short description of it. The idea is old. Homer Dudley of Bell Telephone Laboratories filed a patent for it in 1935, granted as US patent 2,151,091, "Signal transmission" (read 16 November 2026). It describes a system in which "a speech signal is analyzed for its fundamental frequency and the average power in properly chosen subbands of frequency". That small set of numbers is sent instead of the whole sound, and the far end uses it to build a copy of the voice.
The reason was bandwidth. The patent says much of what a speech signal carries is predictable, so it can be made again at the receiving end instead of sent, and "the frequency band width of transmission can be reduced." Pitch tells the machine how high the voice is, and the energy in each band tells it how the sound is shaped.
How does a neural vocoder work in text to speech?
In modern text to speech, a neural vocoder is the last step: it turns a compact picture of the sound into the actual audio. Many systems split the work in two. A first model turns the text into a mel-spectrogram, which in plain words is a chart of how loud each band of pitch is, moment by moment. The vocoder then turns that chart into the waveform you hear.
Two papers show the range. WaveNet (read 16 November 2026) is "a deep neural network for generating raw audio waveforms", and it is autoregressive: "the predictive distribution for each audio sample conditioned on all previous ones". Put simply, it writes the sound one tiny step at a time, each step guided by the steps before it. For text to speech, its authors report that listeners rated it as more natural than the best systems then available, in English and Mandarin.
HiFi-GAN (read 16 November 2026) goes for speed. Its authors show it turning mel-spectrograms into audio, including for speakers the model had not heard, and report 22.05 kHz audio generated 167.9 times faster than real time on a single V100 GPU. Those are the authors' test figures, not a measure of any product.
Why does the vocoder matter on a phone call?
The vocoder decides how clean and natural the voice sounds, and how soon audio can start. A poor one can leave a buzz or a metallic edge that makes numbers and names harder to catch. Speed matters too: a caller who hears a pause thinks the line has dropped. For the reply-time side, see voice agent latency.
The vocoder does not decide what the voice says or how it phrases it. Those come from the earlier steps, and from prosody, the pitch and rhythm of the sentence. To compare voices by ear in Indian languages, read our guide to comparing text-to-speech voices in Indian languages.
On Bananaflow, you choose the voice for each agent, and the agent speaks English plus 10 Indian languages. Read how a Bananaflow agent speaks English and 10 Indian languages.
FAQ
Frequently asked questions
What is a vocoder?
A vocoder is a system that breaks speech into a small set of measurements, such as pitch and loudness in each frequency band, and rebuilds sound from them. In text to speech, a neural vocoder writes the final audio waveform.
Who invented the vocoder?
Homer Dudley of Bell Telephone Laboratories. His US patent 2,151,091, filed in 1935 and granted in 1939, describes analysing speech for its pitch and band energies and rebuilding it at the far end, to cut the bandwidth speech needs.
What is a neural vocoder?
A neural vocoder is a trained model that turns a compact description of sound, usually a mel-spectrogram, into an audio waveform. WaveNet and HiFi-GAN are two published examples, and both are used in text-to-speech research.
What is a mel-spectrogram?
A mel-spectrogram is a chart of how loud each band of pitch is at each moment in a recording, with the bands spaced to follow how people hear. It is a compact picture of the sound that a vocoder can turn back into audio.
Is a vocoder the same as text to speech?
No. Text to speech is the whole job of turning written text into spoken audio. In many systems the vocoder is the last step in it, after other steps have worked out the words to say and the sound to make.
Can I choose the voice of a Bananaflow agent?
Yes. You choose the voice for each agent, and the agent speaks English plus 10 Indian languages. Pre-recorded audio can be mixed with generated speech, so a fixed opening line sounds exactly as you recorded it.
