// basically don't process anything that is less than 1.0s
// see issue #39: https://github.com/ggerganov/whisper.cpp/issues/39
if (seek_end < seek_start + (params.speed_up ? 50 : 100)) {
+ WHISPER_PRINT_DEBUG("%s: input is too short - %d ms < 1000 ms\n", __func__, (seek_end - seek_start)*10);
return 0;
}
// do not allow to go back in time
if (has_ts && seek_delta > seek_delta_new && result_len < i) {
+ WHISPER_PRINT_DEBUG("%s: decoder %d: failed due to seek_delta (%d > %d)\n", __func__, j, seek_delta, seek_delta_new);
failed = true; // TODO: maybe this is not a failure ?
continue;
}
if (seek + seek_delta + 100 >= seek_end) {
result_len = i + 1;
} else {
+ WHISPER_PRINT_DEBUG("%s: decoder %d failed (result_len = 0)\n", __func__, j);
failed = true;
continue;
}
seek_delta = 100*WHISPER_CHUNK_SIZE;
}
+ WHISPER_PRINT_DEBUG("%s: decoder %d completed\n", __func__, j);
completed = true;
continue;
}
// sometimes, the decoding can get stuck in a repetition loop
// this is an attempt to mitigate such cases - we flag the decoding as failed and use a fallback strategy
if (i == n_max - 1 && (result_len == 0 || seek_delta < 100*WHISPER_CHUNK_SIZE/2)) {
+ WHISPER_PRINT_DEBUG("%s: decoder %d: failed due to repetition loop\n", __func__, j);
failed = true;
continue;
}
WHISPER_PRINT_DEBUG("%s: best decoder = %d\n", __func__, best_decoder_id);
}
+ bool success = true;
+
// was the decoding successful for the current temperature?
// do fallback only if:
// - we are not at the last temperature
- // - we are not at the end of the audio (3 sec)
- if (it != (int) temperatures.size() - 1 &&
- seek_end - seek > 10*WHISPER_CHUNK_SIZE) {
- bool success = true;
-
+ if (it != (int) temperatures.size() - 1) {
const auto & decoder = state->decoders[best_decoder_id];
if (decoder.failed || decoder.sequence.avg_logprobs < params.logprob_thold) {
+ WHISPER_PRINT_DEBUG("%s: failed due to avg_logprobs %8.5f < %8.5f\n", __func__, decoder.sequence.avg_logprobs, params.logprob_thold);
success = false;
state->n_fail_p++;
}
+ }
- if (success) {
- //for (auto & token : ctx->decoders[best_decoder_id].sequence.tokens) {
- // WHISPER_PRINT_DEBUG("%s: token = %d, p = %6.3f, pt = %6.3f, ts = %s, str = %s\n", __func__, token.id, token.p, token.pt, ctx->vocab.id_to_token.at(token.tid).c_str(), ctx->vocab.id_to_token.at(token.id).c_str());
- //}
+ if (success) {
+ //for (auto & token : ctx->decoders[best_decoder_id].sequence.tokens) {
+ // WHISPER_PRINT_DEBUG("%s: token = %d, p = %6.3f, pt = %6.3f, ts = %s, str = %s\n", __func__, token.id, token.p, token.pt, ctx->vocab.id_to_token.at(token.tid).c_str(), ctx->vocab.id_to_token.at(token.id).c_str());
+ //}
- break;
- }
+ break;
}
WHISPER_PRINT_DEBUG("\n%s: failed to decode with temperature = %.2f\n", __func__, t_cur);