openai / whisper-medium

## Overview Whisper Medium is an automatic speech recognition model from OpenAI. It splits audio into 30-second chunks, converts each chunk into a log-Mel spectrogram, then decodes text with an encoder-decoder Transformer. It produces readable transcripts and time-aligned segments, so you can search, caption, and index audio at scale. OpenAI released Whisper weights and code under the Apache-2.0 license. ## What you can build - Meeting notes from recordings, with speaker-labeled segments when you enable diarization - Podcast and lecture transcripts with timestamped segments for chaptering - Call center QA transcripts for review and compliance workflows - Subtitle drafts for video editors that need time anchors - Searchable archives for interviews, research sessions, and usability tests ## Inputs - One audio file you upload as WAV, M4A, MP3, OGG, OPUS, or WebM. - The spoken language selection. You can pick a language or let the system auto-detect it. - A limit on how much text the model may generate per chunk. Use this to prevent runaway output. - The chunk length in seconds for long recordings. Shorter chunks reduce context per decode. - A batch size for how many chunks process together. Lower this if you hit memory limits. - An optional speaker diarization toggle to split speech by speaker turns. - An optional expected speaker count. Provide it when you know the exact number of speakers. ## Outputs The model returns a structured transcription result. It includes the full transcript text plus a list of timestamped segments. Each segment typically includes a start time, end time, and the text spoken in that interval. If you enable diarization, segments include speaker labels so you can attribute text to each speaker. The response may also include a detected language label when you use auto language detection. ## Recommended settings - Keep chunk length near 30 seconds for general speech. Whisper was designed around 30-second windows. - Set the language explicitly when you know it. This can improve accuracy on noisy audio. - Turn on diarization for interviews and meetings. Set the speaker count when the participant count is fixed. - Reduce batch size when transcribing long files on limited hardware. This lowers peak memory use. ## Limitations - The model can hallucinate text that was not spoken, especially with noise or low-resource languages. - Accuracy varies by language, accent, dialect, and recording conditions. - The model may repeat phrases or get stuck in loops on some inputs. - Chunking can split words or sentences at boundaries. This can affect punctuation and timing. - Overlapping speakers and loud background music reduce both transcription and diarization quality. - Low-quality inputs raise error rates. Examples include heavy compression, clipping, room echo, or inconsistent volume. ## Safety & compliance - Get consent before transcribing private conversations or recordings of individuals. - Don’t use transcripts for high-stakes decisions without human review. - Don’t use outputs to infer sensitive traits about speakers. - Follow applicable privacy, wiretapping, and data retention laws for your jurisdiction.