baidu-research/warp-ctc

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

4069 stars · 1025 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

A fast parallel implementation of the Connectionist Temporal Classification (CTC) loss function for CPU and CUDA GPU, with a simple C interface and Torch bindings. It is designed for training sequence models such as end-to-end speech recognition systems without needing input-label alignments.

Use cases

• compute CTC loss for training speech recognition models
• train end-to-end sequence models without frame-level alignments
• speed up CTC computation on GPU for large-scale training
• integrate a fast CTC loss into a custom deep learning framework via the C API
• keep training data on GPU to maximize data parallelism

When to choose

• you need a numerically stable, high-performance CTC loss on CPU or NVIDIA GPU
• you are training recurrent or end-to-end speech recognition models at scale
• you want a C interface to embed CTC into your own training stack

When to avoid

• you need bindings for modern frameworks like PyTorch or TensorFlow (built-in CTC may suffice)
• your project requires actively maintained dependencies or recent GPU architectures
• you are not training sequence models that need CTC

Facets

library · maturity maintenance

machine-learning deep-learning gpu-computing deep-learning speech-processing machine-learning windows cpp ctc-loss speech-recognition sequence-learning torch-bindings cuda linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="baidu-research/warp-ctc")