extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-12-26 16:48:50 +01:00
Code Issues Packages Projects Releases Wiki Activity

whisper : suppress non-speech-related token outputs (#473)

Browse Source 
* add non-speech-token suppression

* add suppress non-speech_tokens param
This commit is contained in:
shibukazu 2023-02-08 16:05:34 +09:00 committed by GitHub
parent 2bfe0ebc0f
commit cfc06bf8df
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 37 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
whisper.cpp
View File

@ -2936,6 +2936,7 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
/*.language =*/ "en",
/*.suppress_blank =*/ true,
/*.suppress_non_speech_tokens =*/true,
/*.temperature =*/ 0.0f,
/*.max_initial_ts =*/ 1.0f,
@ -3077,6 +3078,14 @@ static int whisper_wrap_segment(struct whisper_context & ctx, int max_len, bool
return res;
}
static const std::vector<std::string> non_speech_tokens
{
"\"", "#", "(", ")", "*", "+", "/", ":", ";", "<", "=", ">", "@", "[", "\\", "]", "^",
"_", "`", "{", "|", "}", "~", "「", "」", "『", "』", "<<", ">>", "<<<", ">>>", "--",
"---", "-(", "-[", "('", "(\"", "((", "))", "(((", ")))", "[[", "]]", "{{", "}}", "♪♪",
"♪♪♪","♩", "♪", "♫", "♬", "♭", "♮", "♯"
};
// process the logits for the selected decoder
// - applies logit filters
// - computes logprobs and probs
@ -3137,6 +3146,33 @@ static void whisper_process_logits(
logits[vocab.token_translate] = -INFINITY;
logits[vocab.token_transcribe] = -INFINITY;
// suppress non-speech tokens
// ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
if (params.suppress_non_speech_tokens)
{
for (const std::string &token : non_speech_tokens)
{
std::string suppress_tokens[] = {token, " " + token};
for (const std::string &suppress_token : suppress_tokens)
{
if (vocab.token_to_id.find(suppress_token) != vocab.token_to_id.end())
{
logits[vocab.token_to_id.at(suppress_token)] = -INFINITY;
}
}
}
// allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
if (vocab.token_to_id.find(" -") != vocab.token_to_id.end())
{
logits[vocab.token_to_id.at(" -")] = -INFINITY;
}
if (vocab.token_to_id.find(" '") != vocab.token_to_id.end())
{
logits[vocab.token_to_id.at(" '")] = -INFINITY;
}
}
// timestamps have to appear in pairs, except directly before EOT; mask logits accordingly
// https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L414-L424
{

1
whisper.h
View File

@ -285,6 +285,7 @@ extern "C" {
// common decoding parameters:
bool suppress_blank; // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
bool suppress_non_speech_tokens; // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
float temperature; // initial decoding temperature, ref: https://ai.stackexchange.com/a/32478
float max_initial_ts; // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L97