PAPERS-06-005 · Paper — Tacotron

The paper

Tacotron, introduced by Wang et al. in 2017 at Interspeech, is a landmark paper because it showed that a single neural network could turn text into speech, removing the need for many handcrafted components.

Before this paper

Before Tacotron, text‑to‑speech systems were built as pipelines with separate hand‑engineered stages: text analysis, duration modeling, and acoustic modeling. Each stage required expert knowledge and careful alignment, making development slow and brittle. The field struggled to reduce this complexity while keeping quality high. Researchers sought a simpler, more unified approach.

The key idea

The key insight of Tacotron is to map characters directly to a spectrogram using one sequence‑to‑sequence network with attention, training the whole system end‑to‑end from paired text and audio. This eliminates the need for separate modules and hand‑crafted features. By learning the alignment internally, the model can generate natural‑sounding speech with far less engineering effort.

How it works

First, an encoder reads the input characters and turns them into a series of hidden vectors. An attention mechanism then learns how to align each character representation with the appropriate output time steps. The decoder, guided by this attention, generates a spectrogram frame by frame. Each generated frame is conditioned on previously produced frames, allowing smooth continuity. Finally, a vocoder (in the original Tacotron a Griffin‑Lim algorithm, later paired with a WaveNet‑style vocoder in Tacotron 2) converts the spectrogram into an audible waveform. The whole pipeline is trained jointly on text‑audio pairs, so the network learns both pronunciation and timing.

See it — text straight to spectrogram

flowchart LR
  A["Characters"] --> B["Seq2seq<br/>+ attention"]
  B --> C["Spectrogram"]
  C --> D["Vocoder"]
  D --> E["Waveform"]

What they showed

The authors demonstrated that Tacotron could produce natural speech from text with far less hand engineering than traditional pipelines, achieving high‑quality results without the complex multi‑stage setups used before.

What it changed

Tacotron changed the field by making end‑to‑end neural text‑to‑speech the standard approach. It showed that a single model could replace the entire handcrafted pipeline, inspiring many follow‑up works and newer architectures. This shift dramatically simplified TTS development and accelerated progress toward human‑like synthetic voices. As a result, modern TTS systems today trace their design back to Tacotron’s ideas. Next: wav2vec 2.0.

The one-line takeaway

Tacotron proved that a single attention‑based seq2seq model can replace the whole traditional TTS pipeline.

Read it / meet it

Reading the original Tacotron paper gives insight into how the authors built the first successful end‑to‑end TTS system. Its legacy lives on in today’s high‑quality neural voices, many of which still use the encoder‑attention‑decoder framework introduced by Tacotron.