I've processed batch transcription jobs that ran 14 hours overnight and came back with 40% of the files rejected because someone skipped audio preprocessing. Batch mode is where speech-to-text APIs earn their keep. You submit hundreds of files, walk away, and collect transcripts in the morning. It's also where small mistakes compound into expensive reruns. In this guide, I'll walk through the batch transcription API patterns I've used across seven production deployments: preprocessing with FFmpeg, retry logic that doesn't double your bill, and cost controls that stick.
In our complete guide to speech-to-text for AI agents, we covered how voice pipelines fit together. This piece focuses on the batch path and what you send before you worry about delivery.
Why batch transcription still wins for most workloads
Batch transcription means you submit audio files asynchronously and collect results when processing finishes. No persistent connection, no streaming buffer management, no real-time latency budget. For agent workloads that process recorded calls, uploaded voicemails, or podcast episodes, batch is usually the right default.
I've benchmarked both paths on a 500-hour dataset. Batch through Privocio's async endpoint averaged $0.05/hour on our Go plan. Real-time streaming on Deepgram hit $0.36/hour for the same audio.
Batch jobs finish in 1-5 minutes for a 30-minute file. If your agent needs sub-second response, use streaming. For recorded call analysis or training data extraction, batch wins.
For webhook-based async delivery, see our async transcription with webhooks guide.
Audio preprocessing that cuts failures in half
Most batch failures I've debugged trace back to the input file, not the API. Corrupt headers, wrong sample rates, stereo channels with one side silent, or files encoded at 8kHz when the model expects 16kHz.
Here's the FFmpeg pipeline I run on every file before upload:
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
That single command fixes 80% of rejection issues I've seen. It normalizes to 16kHz mono PCM, the format OpenAI Whisper and most Whisper-based APIs expect.
I also check duration, silence ratio, and bitrate before upload.
| Input issue | Symptom | Fix |
|---|---|---|
| Wrong sample rate | Garbled or empty transcript | Resample to 16kHz with FFmpeg |
| Stereo with one silent channel | Half-speed or missing words | Convert to mono (-ac 1) |
| Compressed low-bitrate MP3 | Low word error rate scores | Re-encode to WAV before upload |
| Files over 25MB | Upload timeout or 413 error | Split into 10-minute chunks |
| Leading/trailing silence | Wasted processing time | Trim with silenceremove filter |
Run preprocessing in your upload pipeline, not as a manual step.
Chunking long files without breaking context
Most transcription APIs cap file size somewhere between 25MB and 100MB. A 2-hour meeting recording at reasonable quality blows past that limit fast.
Split at natural silence boundaries, never at arbitrary byte offsets.
ffmpeg -i long_recording.mp3 -f segment -segment_time 600 -reset_timestamps 1 chunk_%03d.wav
Stitch transcripts with a 2-second overlap between chunks and deduplicate. Use Privocio's Agent output mode on each chunk for cleaner stitching.
Retry patterns that don't double your bill
Batch APIs retry failed jobs internally. Your application should retry too, but with guardrails.
The pattern I use: exponential backoff starting at 5 seconds (cap at 5 minutes), max 3 application-level retries, idempotency by file hash (hash the preprocessed file, not the original), and separate handling for transient errors (429, 503) versus permanent ones (400, 422).
A 400 means your file is bad. Retrying a corrupt file three times means you pay for three failed processing attempts on usage-based APIs. On Privocio's fixed pricing, you still waste queue time and delay downstream pipelines.
Log every retry with the job ID, error code, and file hash.
Cost control at scale
Batch transcription costs scale with hours processed, not API calls. Three levers move the number:
- Preprocess to reduce duration — trimming silence on call recordings typically cuts billable minutes by 15-25%
- Pick the right output mode — Agent mode produces shorter transcripts that cost less when you feed them to an LLM downstream. I've measured 35-45% token reduction versus raw output on support call datasets
- Batch during off-peak — if your provider offers queue priority tiers, standard queue is fine for overnight jobs
At 400 hours per month, Privocio's Go plan covers everything for $19/4 weeks. Test your pipeline on the free tier before committing.
Frequently Asked Questions
What audio format should I send to a batch transcription API?
Send 16kHz mono WAV or FLAC if you control encoding. If you're uploading MP3 or M4A, run FFmpeg to normalize first. I've seen accuracy drop 8-12% on heavily compressed inputs compared to clean WAV.
How long does batch transcription take?
For a 30-minute file, expect 1-5 minutes on most Whisper-based APIs. Longer files scale roughly linearly. A 2-hour recording chunked into twelve 10-minute segments processes in parallel if your API supports concurrent jobs.
Should I use webhooks or polling for batch jobs?
Webhooks. Polling works until you have 200+ concurrent jobs, then your workers spend more time checking status than processing results. Our webhook guide covers the handler patterns.
How do I handle files that fail transcription twice?
Dead-letter them after two failures and flag for manual review. Files that fail twice usually have audio quality issues that no retry will fix. Check the file with ffprobe and listen to a 30-second sample.
Does batch transcription work for real-time agent use cases?
No. Batch is for recorded audio processed after the fact. If your agent needs live voice input, you need streaming transcription. For recorded call analysis, meeting notes, and training data extraction, batch is the right choice.
Conclusion: preprocess before you scale
Batch transcription is simple until it isn't. The teams that scale cleanly preprocess every file, chunk at silence boundaries, retry with idempotency, and pick output modes that match their downstream cost profile. The teams that struggle skip preprocessing and wonder why 40% of their jobs fail.
If you're building a batch pipeline for agent workloads, start with FFmpeg normalization and test on the free tier. For the full voice pipeline picture, read our speech-to-text for AI agents guide. When you're ready to run production volume, pricing starts at $19/4 weeks with no per-minute surprises.
Image Credits:
Cover image sourced from Unsplash (Unsplash License).