Current and historical sources are version-scoped. These are original structured annotations, grounded in the linked public upstream files. No GPU benchmark or inference run was performed.
Status: current. Conditions: {"device":"cuda","cuda":12,"cudnn":9}
The current faster-whisper README requires cuBLAS for CUDA 12 and cuDNN 9 for CUDA 12.
Action: Use the current documented GPU stack; confirm supported compute types on the actual GPU.
Evidence literal: cuDNN 9 for CUDA 12
Original sourceStatus: current. Conditions: {"device":"cuda","cuda":11,"cudnn":8}
The current README documents CTranslate2 3.24.0 as a workaround for CUDA 11 and cuDNN 8.
Action: Pin ctranslate2==3.24.0 in an isolated environment; verify the full dependency set before deployment.
Evidence literal: 3.24.0
Original sourceStatus: current. Conditions: {"device":"cuda","cuda":12,"cudnn":8}
The current README documents CTranslate2 4.4.0 as a workaround for CUDA 12 and cuDNN 8.
Action: Pin ctranslate2==4.4.0 in an isolated environment; verify dependencies before deployment.
Evidence literal: 4.4.0
Original sourceStatus: historical. Conditions: {"device":"cuda"}
The v1.0.0 README lists cuBLAS for CUDA 11 and cuDNN 8 for CUDA 11. This is historical documentation, not a current installation instruction.
Action: Read this only when reproducing the v1.0.0 release. Do not replace a current rule with this snapshot.
Evidence literal: cuDNN 8 for CUDA 11
Original sourceStatus: current. Conditions: {"input":"array"}
transcribe decodes file inputs but passes NumPy arrays straight to a feature extractor configured for a 16000 Hz sampling rate.
Action: Decode PCM using the correct sample format, normalize to float32, convert channels to mono, and resample to 16000 Hz before array input.
Evidence literal: if not isinstance(audio, np.ndarray):
Original sourceStatus: current. Conditions: {"input":"file"}
The README says PyAV bundles FFmpeg libraries, so a separate system FFmpeg installation is not required for file decoding.
Action: Use a supported file path for built-in decoding. Do not transfer this claim to raw NumPy arrays.
Evidence literal: FFmpeg does **not** need to be installed
Original sourceStatus: current. Conditions: {}
Transcription starts when the segments generator is iterated.
Action: Iterate over segments or call list(segments) before measuring completed transcription.
Evidence literal: segments = list(segments)
Original sourceStatus: current. Conditions: {"device":"cpu"}
The current README demonstrates CPU inference with compute_type="int8".
Action: Use the documented CPU example as a starting point; measure on your own hardware.
Evidence literal: device="cpu", compute_type="int8"
Original source