Summary
Kyoai has released the full training stack for Pocket TTS, a lightweight text-to-speech model that runs on CPU at inference time without requiring a GPU. Fahd Mirza walks through the complete end-to-end training process on an Ubuntu machine equipped with an H100, covering all four pipeline stages: data preparation using LibriVox audio paired with word-level alignment from the QI Align dataset, forced alignment mapping words to timestamps, training a 24-layer teacher model on GPU, and distilling it down to a 6-layer student model optimized for CPU deployment.
The training pipeline installs via a single UV sync command and uses JSONL manifest files tracking audio paths, start times, durations, and word-level alignments. Quality milestones emerge at recognizable intervals: around 15,000 training steps intelligible words appear, by 50,000 steps the model reads arbitrary text, and by 200,000 steps the voice loses its synthetic quality. Mirza demonstrates how to substitute any custom dataset — whether a different language, a target speaker’s voice, or a proprietary corpus — by simply replacing the source audio and transcript files before the alignment step.
This walkthrough is particularly relevant for developers building voice applications on edge devices, embedded systems, or offline environments where GPU-free inference is a hard requirement. The full training repository is open-source, and Mirza shows that even commodity GPU hardware is sufficient for the training phase.
📺 Source: Fahd Mirza · Published August 30, 2026
🏷️ Format: Hands On Build
