Augmentation is currently MUSAN additive noise only (SNR 20/15 dB for the Whisper runs, 20/13 dB plus music for the LFM runs). Table 6 shows it is doing a lot of work — without it CS-WER sits at the untrained baseline — but it only varies one axis: additive background noise.
The mismatches that actually hurt on real speech are channel, speaker, and rate, none of which MUSAN addresses.
What's needed
- Room impulse response convolution — reverberation is the biggest untested channel mismatch; meeting rooms are reverberant and our corpus is anechoic
- Speed perturbation at 0.9x/1.1x (Ko et al. 2015) — cheap speaker-rate variation
- SpecAugment (Park et al. 2019) — time/frequency masking, complementary to additive noise
- Pitch shifting / vocal tract length perturbation — partial mitigation for the single-voice-per-language problem
- More TTS voices — the cheapest fix for single-voice memorisation; Azure has several ja-JP and en-US neural voices we are not using
Priority
More TTS voices and RIR first — they attack the two failure modes §8 identifies as most likely responsible for the synthetic-to-real gap.
Done when
An ablation table like Table 6(a) but over augmentation type rather than SNR level, scored on the real benchmark rather than the synthetic one.
Augmentation is currently MUSAN additive noise only (SNR 20/15 dB for the Whisper runs, 20/13 dB plus music for the LFM runs). Table 6 shows it is doing a lot of work — without it CS-WER sits at the untrained baseline — but it only varies one axis: additive background noise.
The mismatches that actually hurt on real speech are channel, speaker, and rate, none of which MUSAN addresses.
What's needed
Priority
More TTS voices and RIR first — they attack the two failure modes §8 identifies as most likely responsible for the synthetic-to-real gap.
Done when
An ablation table like Table 6(a) but over augmentation type rather than SNR level, scored on the real benchmark rather than the synthetic one.