引導式測驗 · 中等 等級
用於本地轉錄測驗的 Faster-Whisper 和 Whisper.cpp
比較本地 Whisper 運行時並評估模型大小、量化、硬體、轉錄品質和隱私責任。
問題 1 的 8
根據專案描述,Faster-Whisper 使用的是哪種推理引擎?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 根據專案描述,Faster-Whisper 使用的是哪種推理引擎?
- PyTorch
- whisper.cpp
- CTranslate2
- ONNX Runtime
- How is whisper.cpp described by its project maintainers?
- A Python implementation using CTranslate2
- A C and C++ port of Whisper
- A Python client that requires a hosted transcription service
- A model checkpoint with no inference runtime
- 量化可以用什麼來換取減少的記憶體或更快的推理?
- Only disk space, with no effect on output quality
- Word error rate is independent of model precision
- Quantization guarantees the same transcript as full precision
- Some transcription quality
- 為什麼運行時之間的原始速度比較經常會產生誤導?
- Model, device, decoding settings and workload may differ
- The model size and precision are matched
- The same input duration and device are used
- Decoding and batching settings are equivalent
- 有意義的評估需要什麼字錯誤率?
- Reference transcripts with unstated text normalization
- Only elapsed inference time
- Reference transcripts and stated normalization conventions
- The number of model parameters without any transcript
- 本地推理是否會自動保證隱私合規性?
- Yes; local execution removes every consent obligation
- No; storage, logs, access and consent still matter
- Yes; local model weights guarantee all stored recordings are protected
- Yes; downloading a model once prevents all later content logging
- A deployment transcribes long, noisy meetings in a particular language. Which benchmark best supports decisions about that deployment?
- Short clean clips in a different language
- Silence-only clips used to isolate runtime overhead
- Studio speech without the overlaps or noise expected in use
- Representative recordings from intended conditions
- When should a human review a locally generated transcript?
- When errors could affect records, decisions or quotations
- Only when audio was also uploaded to a hosted service
- Only when the runtime reports an execution failure
- Only when a nondefault checkpoint was selected