TorchCodec 0.14: HDR Video Decoding for CPU and CUDA, and Fast Wav Decoder

Meta's PyTorch team released TorchCodec 0.14, adding a fast WavDecoder that bypasses FFmpeg for direct WAV file reading and HDR video decoding support for CPU and CUDA with full float32 precision. The update also improves audio seeking speed and removes the NVIDIA NPP dependency for simpler CUDA decoding setup.

TorchCodec 0.14 https://meta-pytorch.org/torchcodec/0.14/ is out It is compatible with torch = 2.11 . It comes with two major additions: a fast audio WavDecoder https://meta-pytorch.org/torchcodec/0.14/generated/torchcodec.decoders.WavDecoder.html torchcodec.decoders.WavDecoder , and support for HDR video decoding https://meta-pytorch.org/torchcodec/0.14/generated examples/decoding/hdr decoding.html Fast wav decoder TorchCodec now has a dedicated WavDecoder https://meta-pytorch.org/torchcodec/0.14/generated/torchcodec.decoders.WavDecoder.html torchcodec.decoders.WavDecoder for decoding WAV files. It bypasses FFmpeg entirely and reads WAV data directly, resulting in significantly faster decoding. It supports multiple sample formats int16, int32, float32, etc. , and can decode from files, bytes, or file-like objects. python from torchcodec.decoders import WavDecoder decoder = WavDecoder "audio.wav" samples = decoder.get all samples AudioSamples with data and sample rate HDR Video Decoding VideoDecoder now supports HDR https://meta-pytorch.org/torchcodec/0.14/generated examples/decoding/hdr decoding.html High Dynamic Range video decoding without losing precision. When output dtype=torch.float32 is specified, the decoder outputs RGB float32 frames in 0, 1 , preserving the full HDR color range. This is supported for both CPU and CUDA python import torch from torchcodec.decoders import VideoDecoder decoder = VideoDecoder "hdr video.mp4", output dtype=torch.float32 frame = decoder 0 Full HDR precision in float32 Other Improvements Improved audio seeking :seeking is now much faster AudioDecoder 1449 https://github.com/meta-pytorch/torchcodec/pull/1449 Dropped NPP dependency : TorchCodec no longer depends on NVIDIA's NPP library, which will simplify installing and using TorchCodec for CUDA decoding.