From: Georgi Gerganov Date: Sat, 4 Feb 2023 07:03:14 +0000 (+0200) Subject: whisper : suppress task tokens (#442) X-Git-Tag: upstream/1.7.4~1586 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=291980369c1d3cba5d85f78b77b8f944b821d1cc;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp whisper : suppress task tokens (#442) --- diff --git a/whisper.cpp b/whisper.cpp index d52143f2..97ec46e0 100644 --- a/whisper.cpp +++ b/whisper.cpp @@ -2956,6 +2956,10 @@ static void whisper_process_logits( logits[vocab.token_sot] = -INFINITY; logits[vocab.token_solm] = -INFINITY; + // suppress task tokens + logits[vocab.token_translate] = -INFINITY; + logits[vocab.token_transcribe] = -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 {