#14 Add option to pass on float entropy_thold, logprob_thold, beam_size, best_of, split_on_word when doing the prediction

Author: jwijffels

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: audio.whisper
 Type: Package
 Title: Transcribe Audio Files using the "Whisper" Automatic Speech Recognition Model
-Version: 0.2.1-1
+Version: 0.2.2
 Maintainer: Jan Wijffels <jwijffels@bnosac.be>
 Authors@R: c(
     person('Jan', 'Wijffels', role = c('aut', 'cre', 'cph'), email = 'jwijffels@bnosac.be', comment = "R wrapper"), 

NEWS.md

@@ -1,3 +1,7 @@
+## CHANGES IN audio.whisper VERSION 0.2.2
+
+- Add option to pass on float entropy_thold, logprob_thold, beam_size, best_of, split_on_word when doing the prediction
+
 ## CHANGES IN audio.whisper VERSION 0.2.1-1
 
 - whisper_download_model now Deprecates downloading from https://ggml.ggerganov.com and changed the URL's to download models from huggingface (Issue #18)

R/RcppExports.R

@@ -5,8 +5,8 @@ whisper_load_model <- function(model) {
     .Call('_audio_whisper_whisper_load_model', PACKAGE = 'audio.whisper', model)
 }
 
-whisper_encode <- function(model, path, language, token_timestamps = FALSE, translate = FALSE, print_special = FALSE, duration = 0L, offset = 0L, trace = FALSE, n_threads = 1L, n_processors = 1L) {
-    .Call('_audio_whisper_whisper_encode', PACKAGE = 'audio.whisper', model, path, language, token_timestamps, translate, print_special, duration, offset, trace, n_threads, n_processors)
+whisper_encode <- function(model, path, language, token_timestamps = FALSE, translate = FALSE, print_special = FALSE, duration = 0L, offset = 0L, trace = FALSE, n_threads = 1L, n_processors = 1L, entropy_thold = 2.40, logprob_thold = -1.00, beam_size = -1L, best_of = 5L, split_on_word = FALSE) {
+    .Call('_audio_whisper_whisper_encode', PACKAGE = 'audio.whisper', model, path, language, token_timestamps, translate, print_special, duration, offset, trace, n_threads, n_processors, entropy_thold, logprob_thold, beam_size, best_of, split_on_word)
 }
 
 whisper_print_benchmark <- function(model, n_threads = 1L) {

R/whisper.R

@@ -11,7 +11,7 @@
 #' \itemize{
 #' \item{n_segments: the number of audio segments}
 #' \item{data: a data.frame with the transcription with columns segment, text, from and to}
-#' \item{tokens: a data.frame with the transcription tokens with columns segment, token, token_prob indicating the token probability given the context}
+#' \item{tokens: a data.frame with the transcription tokens with columns segment, token_id, token, token_prob indicating the token probability given the context}
 #' \item{params: a list with parameters used for inference}
 #' }
 #' @export

README.md

@@ -19,7 +19,7 @@ This repository contains an R package which is an Rcpp wrapper around the [whisp
 
 ### Installation
 
-For the *stable* version of this package: `remotes::install_github("bnosac/audio.whisper", ref = "0.2.1-1")` <br>
+For the *stable* version of this package: `remotes::install_github("bnosac/audio.whisper", ref = "0.2.2")` <br>
 Look to the documentation of the functions: `help(package = "audio.whisper")`
 
 - For the *development* version of this package: `remotes::install_github("bnosac/audio.whisper")`

src/RcppExports.cpp

@@ -17,8 +17,8 @@ BEGIN_RCPP
 END_RCPP
 }
 // whisper_encode
-Rcpp::List whisper_encode(SEXP model, std::string path, std::string language, bool token_timestamps, bool translate, bool print_special, int duration, int offset, bool trace, int n_threads, int n_processors);
-RcppExport SEXP _audio_whisper_whisper_encode(SEXP modelSEXP, SEXP pathSEXP, SEXP languageSEXP, SEXP token_timestampsSEXP, SEXP translateSEXP, SEXP print_specialSEXP, SEXP durationSEXP, SEXP offsetSEXP, SEXP traceSEXP, SEXP n_threadsSEXP, SEXP n_processorsSEXP) {
+Rcpp::List whisper_encode(SEXP model, std::string path, std::string language, bool token_timestamps, bool translate, bool print_special, int duration, int offset, bool trace, int n_threads, int n_processors, float entropy_thold, float logprob_thold, int beam_size, int best_of, bool split_on_word);
+RcppExport SEXP _audio_whisper_whisper_encode(SEXP modelSEXP, SEXP pathSEXP, SEXP languageSEXP, SEXP token_timestampsSEXP, SEXP translateSEXP, SEXP print_specialSEXP, SEXP durationSEXP, SEXP offsetSEXP, SEXP traceSEXP, SEXP n_threadsSEXP, SEXP n_processorsSEXP, SEXP entropy_tholdSEXP, SEXP logprob_tholdSEXP, SEXP beam_sizeSEXP, SEXP best_ofSEXP, SEXP split_on_wordSEXP) {
 BEGIN_RCPP
     Rcpp::RObject rcpp_result_gen;
     Rcpp::RNGScope rcpp_rngScope_gen;
@@ -33,7 +33,12 @@ BEGIN_RCPP
     Rcpp::traits::input_parameter< bool >::type trace(traceSEXP);
     Rcpp::traits::input_parameter< int >::type n_threads(n_threadsSEXP);
     Rcpp::traits::input_parameter< int >::type n_processors(n_processorsSEXP);
-    rcpp_result_gen = Rcpp::wrap(whisper_encode(model, path, language, token_timestamps, translate, print_special, duration, offset, trace, n_threads, n_processors));
+    Rcpp::traits::input_parameter< float >::type entropy_thold(entropy_tholdSEXP);
+    Rcpp::traits::input_parameter< float >::type logprob_thold(logprob_tholdSEXP);
+    Rcpp::traits::input_parameter< int >::type beam_size(beam_sizeSEXP);
+    Rcpp::traits::input_parameter< int >::type best_of(best_ofSEXP);
+    Rcpp::traits::input_parameter< bool >::type split_on_word(split_on_wordSEXP);
+    rcpp_result_gen = Rcpp::wrap(whisper_encode(model, path, language, token_timestamps, translate, print_special, duration, offset, trace, n_threads, n_processors, entropy_thold, logprob_thold, beam_size, best_of, split_on_word));
     return rcpp_result_gen;
 END_RCPP
 }
@@ -51,7 +56,7 @@ END_RCPP
 
 static const R_CallMethodDef CallEntries[] = {
     {"_audio_whisper_whisper_load_model", (DL_FUNC) &_audio_whisper_whisper_load_model, 1},
-    {"_audio_whisper_whisper_encode", (DL_FUNC) &_audio_whisper_whisper_encode, 11},
+    {"_audio_whisper_whisper_encode", (DL_FUNC) &_audio_whisper_whisper_encode, 16},
     {"_audio_whisper_whisper_print_benchmark", (DL_FUNC) &_audio_whisper_whisper_print_benchmark, 2},
     {NULL, NULL, 0}
 };

src/rcpp_whisper.cpp

@@ -220,7 +220,12 @@ SEXP whisper_load_model(std::string model) {
 // [[Rcpp::export]]
 Rcpp::List whisper_encode(SEXP model, std::string path, std::string language, 
                           bool token_timestamps = false, bool translate = false, bool print_special = false, int duration = 0, int offset = 0, bool trace = false,
-                          int n_threads = 1, int n_processors = 1) {
+                          int n_threads = 1, int n_processors = 1,
+                          float entropy_thold = 2.40,
+                          float logprob_thold = -1.00,
+                          int beam_size = -1,
+                          int best_of = 5,
+                          bool split_on_word = false) {
     whisper_params params;
     params.language = language;
     //params.model = model;
@@ -232,6 +237,13 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
     params.n_threads = n_threads;
     params.n_processors = n_processors;
 
+    params.entropy_thold = entropy_thold;
+    params.logprob_thold = logprob_thold;
+    params.beam_size = beam_size;
+    params.best_of = best_of;
+    params.split_on_word = split_on_word;
+    //params.no_speech_thold = no_speech_thold;
+    //params.temperature_inc = temperature_inc;
 
     //std::string language  = "en";
     //std::string model     = "models/ggml-base.en.bin";
@@ -417,6 +429,7 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
     Rcpp::StringVector transcriptions_from(n_segments);
     Rcpp::StringVector transcriptions_to(n_segments);
     std::vector<int> token_segment_nr;
+    std::vector<int> token_segment_id;
     std::vector<std::string> token_segment_text;
     std::vector<float> token_segment_probability;
     std::vector<std::string> token_segment_from;
@@ -439,7 +452,9 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
             }
             const char * text = whisper_full_get_token_text(ctx, i, j);
             const float  p    = whisper_full_get_token_p   (ctx, i, j);
+            const int tokenid = whisper_full_get_token_id  (ctx, i, j);
             token_segment_nr.push_back(i + 1);
+            token_segment_id.push_back(tokenid);
             std::string str(text);
             token_segment_text.push_back(str);
             token_segment_probability.push_back(p);
@@ -456,6 +471,7 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
     if(token_timestamps){
         tokens = Rcpp::DataFrame::create(
             Rcpp::Named("segment") = token_segment_nr, 
+            Rcpp::Named("token_id") = token_segment_id, 
             Rcpp::Named("token") = token_segment_text, 
             Rcpp::Named("token_prob") = token_segment_probability,
             Rcpp::Named("token_from") = token_segment_from,
@@ -464,6 +480,7 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
     }else{
         tokens = Rcpp::DataFrame::create(
             Rcpp::Named("segment") = token_segment_nr, 
+            Rcpp::Named("token_id") = token_segment_id, 
             Rcpp::Named("token") = token_segment_text, 
             Rcpp::Named("token_prob") = token_segment_probability,
             Rcpp::Named("stringsAsFactors") = false);
@@ -485,7 +502,15 @@ Rcpp::List whisper_encode(SEXP model, std::string path, std::string language,
                                                Rcpp::Named("duration") = duration,
                                                Rcpp::Named("translate") = params.translate,
                                                Rcpp::Named("token_timestamps") = token_timestamps,
-                                               Rcpp::Named("word_threshold") = params.word_thold));
+                                               Rcpp::Named("word_threshold") = params.word_thold,
+                                               Rcpp::Named("entropy_thold") = params.entropy_thold,
+                                               Rcpp::Named("logprob_thold") = params.logprob_thold,
+                                               Rcpp::Named("beam_size") = params.beam_size,
+                                               Rcpp::Named("best_of") = params.best_of,
+                                               Rcpp::Named("split_on_word") = params.split_on_word));
+    
+
+    
     return output;
 }