Running MFA with speaker adaptation off?

[Prad Kadambi]

Hi there! 

I'm trying to compare performance with and without speaker adaptive training.

Does the -n flag disable speaker adaptation? 

'mfa align -n /path/to/audio /path/to/dict language /path/to/output/alignment_no_SAT'

vs to have speaker adaptation on by default:

'mfa align /path/to/audio /path/to/dict language /path/to/output/alignment_with_SAT'

Also the speaker adaptation calculated is fMLLR correct?

Cheers,

Prad K

[Prad Kadambi]

Additionally I see the following issue in the pretrained_aligner.log

Problem 1: Running `mfa align -n` to run without speaker adaptation fmllr seems to run anyways, while uses_speaker_adaptation=False as expected:

features:
  allow_downsample: true
  allow_upsample: true
  feature_type: mfcc
  fmllr_update_type: full
  frame_shift: 10
  high_frequency: 7800
  low_frequency: 20
  pitch: false
  sample_frequency: 16000
  silence_weight: 0.0
  snip_edges: true
  splice_left_context: 3
  splice_right_context: 3
  use_energy: false
  uses_cmvn: true
  uses_deltas: true
  uses_speaker_adaptation: false
  uses_splices: false
  uses_voiced: false

Problem 2: Running `mfa align --clean` normally to enable speaker adaptation still shows uses_speaker_adaptation=False:

features:
  allow_downsample: true
  allow_upsample: true
  feature_type: mfcc
  fmllr_update_type: full
  frame_shift: 10
  high_frequency: 7800
  low_frequency: 20
  pitch: false
  sample_frequency: 16000
  silence_weight: 0.0
  snip_edges: true
  splice_left_context: 3
  splice_right_context: 3
  use_energy: false
  uses_cmvn: true
  uses_deltas: true
  uses_speaker_adaptation: false
  uses_splices: false
  uses_voiced: false

[Prad Kadambi]

Issue is resolved, I run 'mfa align' to align without SAT, and 'mfa train' to align with SAT