faster-whisper on silence/noise/test tones: language detected as 'nn' (Norwegian Nynorsk) with ~0.25-0.5 probability, and decoding is slower than real speech, so tone-based RTF benchmarks mislead
Benchmark and smoke-test Whisper with real speech, and in production either pass language= or gate on info.language_probability and use vad_filter=True. Non-speech input doesn't produce text, but it costs decode time and gives you a confident-looking wrong language.
One run on a busy shared CPU box (faster-whisper 1.2.1, int8, cpu_threads=4, beam_size=1, 3 s clips plus a 5.9 s Piper TTS sentence), all from the same container: - speech: tiny RTF 0.13, base 0.27, transcribed correctly, language en at 0.99-1.00 - silence: tiny RTF 0.70 with language='en' and 1.44 with auto-detect, which reported 'nn' at 0.25 - white noise, auto-detect: 'nn' at 0.43 (tiny) / 0.50 (base), 0 segments - synthetic formant tones, auto-detect: 'en' at 0.34 / 0.61, 0 segments - any of those non-speech clips with vad_filter=True: 0 segments in about 1% of the audio duration
A voice-pipeline team benchmarked on a synthetic tone WAV and got tiny RTF 1.2-3.3 against base 0.21, so it concluded base is faster than tiny. On real speech the order flipped back (tiny was about 2x faster than base). Treat absolute numbers as host-specific, but don't compare model sizes on audio Whisper can't transcribe.
Quick way to get test speech offline: echo "..." | piper --model en_US-lessac-medium.onnx --output_file s.wav, then ffmpeg -i s.wav -ar 16000 -ac 1 speech.wav.