Whisper API keeps returning empty transcript for videos longer than 30 minutes — stuck in production

API
1

Hey everyone,

Running into a consistent issue with Whisper API and longer recordings and not sure what the right fix is.

My setup right now:

  • Pull the Zoom recording → convert to MP3 with FFmpeg → compress under 25MB → send to Whisper API

Works fine for anything under 20 minutes. The moment I go past 30 minutes the transcript either comes back empty or just cuts off mid-sentence with no error message. Whisper just returns 200 with partial or no content.

Already tried a few things:

Splitting into chunks — works but the speaker attribution gets completely lost between chunks and stitching the context back together is messy.

Lowering the bitrate more — quality drops so much that Whisper starts misidentifying words, especially with any background noise or non-native accents.

Switching to gpt-4o-transcribe — hit the 1500 second limit which is actually worse than Whisper for longer calls.

The real frustration is the entire pipeline assumes you have a small local file. For any real meeting or interview recording that is just not realistic without seriously degrading the audio.

Has anyone figured out a solid approach for this? Ideally something that:

  • Takes the recording URL directly without needing to download and re-encode
  • Handles 60-90 minute recordings reliably
  • Keeps speaker labels intact

Open to completely different approaches if Whisper just isn’t the right tool for this use case.

2

The 25MB limit on the OpenAI Whisper API is the main bottleneck here. Compressing a 90-minute recording to fit that size kills the audio quality (hence the hallucinations or empty transcripts).

Since you’re looking for a production-ready solution that handles 60+ minutes and speaker labels, here are three solid approaches:

  1. Deepgram API: It’s often the go-to for long-form audio. It accepts URLs directly, doesn’t have that strict 25MB limit, and has excellent diarization (speaker labels) that stays consistent throughout the entire recording.

  2. AssemblyAI: Similar to Deepgram, it’s built for long files and handles speaker diarization much better than ‘stitching’ Whisper chunks together.

  3. Self-hosted Whisper (Faster-Whisper): If you have the infra (or use a GPU cloud like RunPod), you can run faster-whisper. You won’t have file size limits, and you can use pyannote-audio for speaker labeling, though it requires more setup.

Sticking with OpenAI’s Whisper for 90-minute files will always feel like a hack because of the chunking/compression trade-off.The 25MB limit on the OpenAI Whisper API is the main bottleneck here. Compressing a 90-minute recording to fit that size kills the audio quality (hence the hallucinations or empty transcripts).

Since you’re looking for a production-ready solution that handles 60+ minutes and speaker labels, here are three solid approaches:

  1. Deepgram API: It’s often the go-to for long-form audio. It accepts URLs directly, doesn’t have that strict 25MB limit, and has excellent diarization (speaker labels) that stays consistent throughout the entire recording.

  2. AssemblyAI: Similar to Deepgram, it’s built for long files and handles speaker diarization much better than ‘stitching’ Whisper chunks together.

  3. Self-hosted Whisper (Faster-Whisper): If you have the infra (or use a GPU cloud like RunPod), you can run faster-whisper. You won’t have file size limits, and you can use pyannote-audio for speaker labeling, though it requires more setup.

Sticking with OpenAI’s Whisper for 90-minute files will always feel like a hack because of the chunking/compression trade-off.