Tacotron 2 without the WaveNet bottleneck

What it does

This is NVIDIA’s PyTorch implementation of the Tacotron 2 text-to-speech model, which converts text into mel spectrograms. It deliberately omits WaveNet, expecting you to pair it with a separate mel-to-audio decoder such as NVIDIA’s WaveGlow or nv-wavenet. It trains on the LJSpeech dataset and ships with pre-trained checkpoints for both the spectrogram predictor and a compatible vocoder.

The interesting bit

The value is mostly engineering hygiene. NVIDIA added distributed multi-GPU training and automatic mixed precision via Apex, turning a research demo into something that converges quickly on their hardware. The decoupled design lets you swap vocoders without retraining the spectrogram network, provided you keep the mel representation identical.

Key highlights

• Pre-trained Tacotron 2 and WaveGlow models available for immediate inference
• Distributed training and automatic mixed precision through NVIDIA Apex
• Warm-start from published checkpoints with dataset-dependent text embeddings ignored by default
• Jupyter notebook included for demo inference
• LJSpeech dataset support out of the box

Caveats

• NVIDIA GPU and CUDA/cuDNN are mandatory; there is no CPU fallback
• Mel-spectrogram representation must match exactly between Tacotron 2 and your mel decoder or synthesis will fail
• Dataset-dependent text embedding layers are ignored by default during warm-start, which may surprise you if you are adapting to new vocabulary

Verdict

Grab it if you need a battle-tested TTS front end on NVIDIA hardware and do not mind wiring up your own vocoder. Look elsewhere if you want a turnkey, end-to-end speech synthesizer—this is only half the pipeline.