edit some comments

whisper.cpp

@@ -372,17 +372,17 @@ struct whisper_vocab {
  std::map<id, token> id_to_token;
 
  // reference: https://github.com/openai/whisper/blob/248b6cb124225dd263bb9bd32d060b6517e067f8/whisper/tokenizer.py#L334-L349
- id token_eot = 50256; // anything before that are normal tokens
- id token_sot = 50257;
+ id token_eot = 50256; // end of text, anything before that are normal tokens
+ id token_sot = 50257; // start of transcript
  // task tokens (used only for multilingual models)
  id token_translate = 50357;
  id token_transcribe = 50358;
  // other special tokens
  id token_solm = 50359; // [TDRZ] used by tinydiarize models to indicate speaker turn
- id token_prev = 50360;
- id token_nosp = 50361;
+ id token_prev = 50360; // TODO: I don't understand the meaning of this token
+ id token_nosp = 50361; // no speech
  id token_not = 50362; // no timestamps
- id token_beg = 50363; // begin timestamps
+ id token_beg = 50363; // the first timestamp token <|0.00|>
 
  bool is_multilingual() const {
  return n_vocab >= 51865;
@@ -1261,23 +1261,23 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
  if (i > vocab.token_beg) {
  word = "[_TT_" + std::to_string(i - vocab.token_beg) + "]";
  } else if (i == vocab.token_eot) {
- word = "[_EOT_]";
+ word = "[_EOT_]"; // <|endoftext|>
  } else if (i == vocab.token_sot) {
- word = "[_SOT_]";
+ word = "[_SOT_]"; // <|startoftranscript|>
  } else if (i == vocab.token_translate) {
  word = "[_TRANSLATE_]";
  } else if (i == vocab.token_transcribe) {
  word = "[_TRANSCRIBE_]";
  } else if (i == vocab.token_solm) {
- word = "[_SOLM_]";
+ word = "[_SOLM_]"; // <|startoflm|>
  } else if (i == vocab.token_prev) {
- word = "[_PREV_]";
+ word = "[_PREV_]"; // <|startofprev|>
  } else if (i == vocab.token_nosp) {
- word = "[_NOSP_]";
+ word = "[_NOSP_]"; // <|nospeech|>
  } else if (i == vocab.token_not) {
- word = "[_NOT_]";
+ word = "[_NOT_]"; // <|notimestamps|>
  } else if (i == vocab.token_beg) {
- word = "[_BEG_]";
+ word = "[_BEG_]"; // the first timestamp token <|0.00|>
  } else if (i > vocab.token_sot && i <= vocab.token_sot + vocab.num_languages()) {
  word = "[_LANG_" + std::string(whisper_lang_str(i - vocab.token_sot - 1)) + "]";
  } else {
@@ -4533,7 +4533,6 @@ static void whisper_no_speech_probs(
  const auto & vocab= ctx.vocab;
  const int n_logits = vocab.id_to_token.size();
  const int logits_offset = decoder.i_batch - (prompt_init.size() - 1);
- printf("[%i, %i, %i]\n", decoder.i_batch, (prompt_init.size() - 1), logits_offset);
  double no_speech_probs = 0.0;
 
  WHISPER_ASSERT(n_logits == ctx.vocab.n_vocab);
@@ -5353,8 +5352,6 @@ int whisper_full_with_state(
 
  whisper_process_logits(*ctx, *state, state->decoders[0], params, t_cur);
 
- printf("Non-speach: %f\n", state->decoders[0].sequence.no_speech_probs);
-
  for (int j = 1; j < n_decoders_cur; ++j) {
  auto & decoder = state->decoders[j];
 

whisper.h

@@ -487,7 +487,7 @@ extern "C" {
  float temperature_inc;
  float entropy_thold; // similar to OpenAI's "compression_ratio_threshold"
  float logprob_thold;
- float no_speech_thold; // probability of silence
+ float no_speech_thold;
 
  struct {
  int best_of; // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264