In my mind, this is analogous to the reasons why evaluation of lossy audio compression codecs
requires human listening tests. Simply running some simplistic signal analysis like SNR (Signal-to-Noise Ratio) completely fails to capture the as-perceived quality of a compression implementation.
To explore that analogy: In the case of lossy audio compression, the compressor deliberately introduces quantization noise into the signal. It does so by running a "psychoacoustic model", which attempts to capture a broad quality of human hearing called auditory masking. There's a number of different kinds of masking[1]: a strong tonal sound creates an "umbrella" across nearby frequencies that can mask quieter noise-like sounds. Similarly there's noise-vs-noise masking, as well as forwards- and backwards- temporal masking. (Yes, backwards. A sound can mask perception of a sound that occurred before it.)
In the audio compression case, we've built an algorithm that attempts to characterize exploitable phenomena of human hearing. These masking characteristics aren't perceived quite the same by human listeners as the model, nor even the same between individual human listeners. Thus the need for human listening tests. These, due to the experimental care and human subjects required, are expensive.
Back to speech transcription. Say the end goal is "how well does a human comprehend this transcribed speech"? (e.g. vs some standard, such as the original speech, vs. the original speech transcribed by a skilled specialist, etc.) The problem starts to look pretty similar. We can cite numerical, word-centric error rates, but that fails to capture how well meaning is preserved and transmitted. Imagine a perverse algorithm that did a perfect transcription, but then dropped or altered words for maximum meaning obfuscation. It might equal or even beat the cited error rates but be much harder to actually comprehend.
[1] https://en.wikipedia.org/wiki/Auditory_masking