return language code with transcription

[qxprakash]

when running a non .en model , and running with flag -l auto , it transcribes the audio into different languages , It would have been nice if it would have detected the langauge of the audio and return it in response along with transcription.

[bradmit]

There is the print segment callback which has the information you're referring to in it. In the application I've integrated whisper cpp into.....

void whisper_print_segment_callback(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data)
...
int langId = whisper_full_lang_id_from_state(state);
...
You can then get the lang code and language full name:
whisper_lang_str(langId)
whisper_lang_str_full(langId)

Or am I missing something?

[qxprakash]

Hi @bradmit sorry I missed one important detail here , I wanted the language code along with the transcrption , in the whisper.cpp server , currently the output of the server is something like this {'text': 'transcription of the audio file'} , I haved something like , {'text': 'transcription of the audio file' , detected_lang: "en"}
is something like this possible ? , if yes which parts of the codebase should I modify to achieve this ? , Thanks !!

[sachaarbonel]

in verbose_json mode there is a dictionary of probabilities

{
  "task": "transcribe",
  "language": "english",
  "text": "This is the transcribed text of the audio file.",
  "language_detection": {
    "probability": 0.982,
    "language_probabilities": {
      "en": 0.982,
      "fr": 0.008,
      "es": 0.005,
      "de": 0.003
    }
  },
  "segments": [
    // ... segments array content ...
  ]
}

See PR that got merged
#3044