Transcription word-level BUG

Just tested the model gemini-3.5-transcribe with a 30-minute audio file, and I can tell ya: THIS MODEL IS NOT STABLE AND SO EXPENSIVE!

Just look at this output:


                        {
                            "word": "the",
                            "startOffset": "9.100s",
                            "endOffset": "9.200s"
                        },
                        {
                            "word": "1800s.",
                            "startOffset": "9.200s",
                            "endOffset": "9.900s"
                        },
                        {
                            "word": "I",
                            "startOffset": "1000s",
                            "endOffset": "1000.100s"
                        },
                        {
                            "word": "would",
                            "startOffset": "1000.700s",
                            "endOffset": "1001.100s"
                        },
                        {
                            "word": "We",
                            "startOffset": "1137.900s",
                            "endOffset": "138.100s"
                        },
                        {
                            "word": "create",
                            "startOffset": "138.100s",
                            "endOffset": "138.600s"
                        },
                        {
                            "word": "the",
                            "startOffset": "138.600s",
                            "endOffset": "138.900s"
                        },

Look at the offsets …

Helo @NeonByteNomad ,

Thanks for flagging, we’re looking into this.