Can we implement similar GOP calculations like Kaldi and handle multi read and missed reads

[mywsat]

Can we implement similar GOP calculations like Kaldi and handle multi read and missed reads

[tabahi]

It's already scoring GOP for each phoneme as "confidence". The missed phonemes are estimated and inserted during post-processing when enforce_all_targets=True. In the output json you will see "is_estimated": true for missed phonemes. Although it doesn't miss many phonemes since it's "forced" (1 to 2% for clean English). If you want be check how much a phoneme is wrongfully forced then "confidence" value below 0.1 would mean there is almost no accoustic evidence for that phoneme.
The "confidence" in json output is similar to GOP in Kaldi.

 "phoneme_ts": [
  {
      "phoneme_id": 29,
      "phoneme_label": "b",
      "ipa_label": "b",
      "start_ms": 33.56833267211914,
      "end_ms": 50.35249710083008,
      "confidence": 0.9970603585243225,   // bad pronounciation = low confidence, it's a measure of acoustic evidence
      "is_estimated": false,  // it will be true if this phoneme was missed
      "target_seq_idx": 0,
      "index": 0
  },

For the whole uttererance if GOP is very bad then "bad_alignment" is set to true if crosses a bad alignement threshold which can be set with PhonemeTimestampAligner(bad_confidence_threshold=0.6)

bournemouth-forced-aligner/examples/samples/audio/109867__timkahn__butterfly.vs.json

Line 25 in 5e951e7

"bad_alignment": false

This is based on the latest changes. Make sure you have the latets version over >1.1.2.