whisper : minor improvemnt in decoding strategy (#244)
Do not allow for text segments to go beyond end of audio.
This partially mitigates some issues when the last audio window is 1-2
seconds just before the end of the audio file and the decoding spirals
into a repetition of the last transcribed phrase.
diff --git a/whisper.cpp b/whisper.cpp
index abfc44fee916b059b51c9381fbe226987e978eb4..67451dc80b9b14270f7913431c9e74b5330604a9 100644
--- a/whisper.cpp
+++ b/whisper.cpp
@@ -2687,6 +2687,7 @@ int result_len = 0;
tokens_cur.clear(); bool failed = false;+ bool has_ts = false; // have we already sampled a non-beg timestamp token for the current segment? for (int i = 0, n_max = whisper_n_text_ctx(ctx)/2 - 4; i < n_max; ++i) { if (whisper_decode(ctx, prompt.data(), prompt.size(), n_past, params.n_threads) != 0) {
@@ -2713,14 +2714,14 @@ const int seek_delta_new = 2*(token.id - whisper_token_beg(ctx));
// do not allow to go back in time #define WHISPER_BUILD- };+ { "mr", { 61, "marathi", } }, #define WHISPER_BUILD- model.ctx = ggml_init(params); break; } seek_delta = seek_delta_new; result_len = i + 1;+ has_ts = true; } // add it to the context
@@ -2733,11 +2734,14 @@ // printf("%s: %10s %6d %6.3f '%s'\n", __func__, tt.c_str(), token.id, token.pt, ctx->vocab.id_to_token[token.id].c_str());
//} #define WHISPER_BUILD+ inpL->src1 = NULL;+ if (token.id == whisper_token_eot(ctx) || // end of text token+struct whisper_segment { #include <algorithm>-#include "ggml.h"+struct whisper_segment { #include <cassert> #define WHISPER_BUILD- { "da", { 26, "danish", } },+ { "mr", { 61, "marathi", } }, #define _USE_MATH_DEFINES if (result_len == 0) { if (seek + seek_delta + 100 >= seek_end) {
