QWE AI Academy Here’s the mistake: you record a 60-minute podcast, upload the MP3 to Otter or Whisper, and get back a transcript where Speaker 2 is labeled as Speaker 1 for half the episode, timestamps are off by 30 seconds, and your guest’s name is transcribed as “Colonel Mustard” instead of “Colin Maas-Ter.”
You blame the AI. Audio’s the problem.
The real workflow works backward: decide what you’re willing to fix manually, choose your audio quality battle, then pick the tool that matches.
The Audio Quality Trap
Every tutorial lists Whisper vs Otter vs Descript specs. None tell you this: poor audio quality is the #1 cause of transcription failure (as of 2026 testing), regardless of which AI you use.
What actually breaks transcription:
Background noise – HVAC hum, traffic, keyboard clicks. AI hears consonants (s, f, th) as noise.
Distance from mic. More than 8 inches away? Voice loses bass, sounds thin. AI accuracy drops.
Overlapping speakers – Two people talking at once. The model may merge speakers, skip the quieter one, or hallucinate words to fill gaps.
Low volume – Speech sits near the noise floor. Proper recording levels: -12dB to -6dB peaks.
Professional transcription testing shows AI transcription accuracy ranges from 50% to 93% depending on audio conditions (as of 2025-2026). Clean studio audio with one speaker? 93%. Phone call with echo and crosstalk? 50%.
The punchline: Whisper trained on 680,000 hours of web audio won’t save your echoy Zoom recording. Fix the source or accept 20 minutes of manual cleanup per episode.
Before you commit to any transcription tool, run a 5-minute test with your actual podcast audio – not a demo file. More than 5-10 errors per minute? Your audio setup is the problem.
Manual podcast transcription takes 4-6 hours per 1 hour of audio for experienced transcriptionists (8-9 hours for beginners, per industry benchmarks). AI cuts this to 20-40 minutes total including cleanup. Not instant, but massive.
Cloud Convenience vs API Control
The tool debate isn’t “which is best” – it’s “do you want a product or a component?”
Cloud convenience: Upload → click → download. Otter.ai and Descript handle everything. You get speaker labels, timestamps, web editor. Trade-off: their pricing, their accuracy, their file format.
Otter’s free tier (as of 2026): 600 minutes/month with real-time transcription, Zoom/Teams integration, support for English/French/Spanish. Built for meetings, not editing. Descript gives you 1 hour/month free, but its power is text-based video editing – delete a word in the transcript, the audio disappears. Paid plans start at $12-24/month (early 2026 pricing).
API control: Code required. OpenAI’s Whisper API costs $0.006/minute ($0.36/hour) as of 2026, runs on your server or theirs. You write the code to handle file uploads, chunking, timestamp reconciliation if you add speaker diarization separately.
For podcasters publishing 4 episodes/month (each 60 min): Otter free tier covers it. For agencies processing 100 hours/month: Whisper API costs $36/month vs Descript Business at $600+/month (12 seats × $50).
Neither is “better.” One is a product (pay for convenience). One is an ingredient (pay with engineering time).
Transcribe a Podcast with Whisper API
No-frills path. You need: Python 3.8+, an OpenAI API key, your podcast MP3.
Step 1: Install dependencies, set up your API key.
pip install openai
export OPENAI_API_KEY='your-key-here'Step 2: Write the transcription script. Whisper API accepts audio files, returns text.
from openai import OpenAI
client = OpenAI()
audio_file = open("podcast_episode.mp3", "rb")
transcript = client.audio.transcriptions.create(
model="whisper-1",
file=audio_file
)
print(transcript.text)Works great. Until it doesn’t – cryptic error, no obvious cause.
Step 3: Handle the 25MB file size trap. Whisper rejects files over 25MB – no warning in the docs (as of 2026), just that error. 60-minute podcast at 128kbps? Right at the edge. Solution: chunk the file with ffmpeg before uploading.
ffmpeg -i podcast.mp3 -f segment -segment_time 600 -c copy chunk_%03d.mp3This splits your audio into 10-minute chunks. Transcribe each, concatenate results. Tedious? Yes. Documented? No. Cloud tools handle this invisibly.
Step 4: Add timestamps if you need them. Pass response_format="verbose_json" for word-level timing.
transcript = client.audio.transcriptions.create(
model="whisper-1",
file=audio_file,
response_format="verbose_json"
)JSON comes back with segments – each has start, end, text. The catch: Whisper doesn’t do speaker diarization natively. You get timestamps for what was said, not who said it.
Speaker Diarization
Speaker labels – “Speaker 1: [text]” vs “Speaker 2: [text]” – are the feature everyone wants. Most tools fumble it.
Descript and Otter include speaker diarization. 2-3 speakers, clear audio? Accuracy hits 96-98% (per 2025-2026 benchmarks). Add a 4th speaker or moderate background noise: drops to 90-94%. Phone quality or 6+ speakers? 85-90%.
Whisper doesn’t include speaker ID. You can bolt it on with a separate tool (pyannote.ai, AssemblyAI), but now you’re solving the “reconciliation problem”: Whisper says a sentence ends at 15.48 seconds, your diarization model says Speaker 2 starts at 15.4 seconds. Which do you trust? Every team writes custom logic to match these timestamps. Most solutions are brittle.
The dirty secret from Azure’s own documentation (as of 2026): “Speaker count estimation is not guaranteed to exceed 2 reliably.” Translation: host, co-host, guest? The AI might label all three as “Speaker 1” and “Speaker 2” randomly.
Remember that reconciliation nightmare? Here’s the fix: record each speaker on a separate track (multitrack recording). Transcribe each track individually, label manually. Tedious but eliminates the guessing game.
Three Edge Cases
Timestamp drift in long podcasts. You transcribe a 90-minute episode. Minute 10: timestamps perfect. Minute 60: transcript is 45 seconds ahead of actual audio. Why? Whisper processes audio in 30-second windows. Silence, pauses, background noise get included in duration calculations. The drift compounds (documented in developer communities as of 2025-2026). Syncing transcripts to video or creating timestamped show notes? You’ll spend an hour fixing this.
No automatic fix. Best mitigation: use Adobe Podcast Enhance (free tier) to strip silence before transcribing. Tightens audio, reduces cumulative drift.
Hallucinations on silence. Whisper sometimes invents text when there’s no speech. A 2025 academic study found 1% of Whisper transcriptions contain entire hallucinated phrases – 38% include violence, false authority, or nonsense. Worse on low-resource languages (anything outside English/Spanish/French).
Can’t predict when it happens. You catch it during manual review or you don’t.
Homophones and phonetic ambiguity. “Their” vs “there,” “two” vs “to.” Whisper uses context to guess, but podcasts are conversational – grammar is loose, context unclear. Automated transcription tools misinterpret homophones constantly. If your podcast discusses technical jargon, brand names, or non-English words, expect 3-5 corrections per minute even with 95% overall accuracy.
Free Options
Apple Podcasts now auto-generates transcripts for new episodes in English, French, Spanish, German (iOS 17.4+, as of 2024-2026). Upload your podcast, Apple transcribes it, listeners see text in-app. Zero cost, zero control.
Can’t edit the transcript. Older episodes get transcribed gradually – no timeline given.
YouTube does the same for video podcasts – auto-captions appear under the video. Accuracy is “decent.” YouTube doesn’t give you an editor. Download the auto-generated subtitle file, fix it in a text editor, re-upload.
Google Recorder (Pixel phones only) transcribes in real-time as you record. Shockingly good for free, but reports of accuracy issues plus syncing problems are common (as of 2025-2026). iPhone users: out of luck.
Hidden cost of free tools: time. Your time is worth $50/hour, you spend 20 minutes per episode cleaning up a free transcript? You’re paying $16.67 per episode anyway.
What This Means for You
If you publish 1-2 podcasts per month, solo or with one guest, clean audio: Start with Otter.ai’s free tier. 600 minutes/month covers you, real-time transcription works well for meetings, web editor is intuitive.
Editing video podcasts, want text-based trimming: Descript’s Creator plan ($24/month as of early 2026) is the only tool that ties transcript edits directly to audio/video cuts. 30 hours of transcription per month is plenty for weekly episodes.
Processing 50+ hours/month, have a developer on your team, need custom workflows: Whisper API at $0.006/min ($0.36/hour, 2026 pricing) scales cheaply. Expect to write 200+ lines of code handling file chunking, timestamp reconciliation, error retries.
Need legally-accurate transcripts (depositions, medical consultations): Don’t use AI alone. Hire human transcriptionists or use a hybrid service (Rev offers AI plus human review). AI transcription is 90-95% accurate, but that last 5-10% includes the errors that matter most.
Next step: Record a 5-minute test clip with your actual podcast setup. Run it through Otter (free), Descript (free trial), Whisper API ($0.03 for 5 minutes). Count the errors in each. Fewer than 5 mistakes in 5 minutes? That’s your tool.
Frequently Asked Questions
Can I transcribe a podcast from Spotify or Apple Podcasts directly?
Not through official tools. But RSS feeds sometimes include audio file URLs that third-party tools can access. For your own podcast, always keep the original MP3/WAV. Some tools like Sonix claim to handle Spotify URLs – works by downloading the public audio stream, not through official API access.
Why does my transcript have the wrong speaker names?
AI assigns generic labels like “Speaker 1” and “Speaker 2” based on voice characteristics – doesn’t perform voice recognition to identify specific people. You manually rename speakers after transcription. Some tools (Otter, Descript) learn to recognize recurring voices over time if you consistently label them. First transcription? Always requires manual intervention.
For podcasts with 4+ speakers or similar-sounding voices, expect speaker labeling errors in 10-15% of segments even after you’ve trained the system. One debugging session with three co-hosts burned through 45 minutes just fixing “Speaker 1” randomly becoming “Speaker 3” mid-sentence.
How long does it actually take to transcribe a 1-hour podcast with AI?
AI processing: 1-5 minutes depending on the tool. Whisper API averages 1-3 minutes for an hour of audio (as of 2026), Otter and Descript similarly fast.
That’s just the automated part. Budget 15-30 minutes of manual cleanup per hour of audio to fix speaker labels, correct homophones, adjust timestamps, catch hallucinations. Poor audio quality (echo, background noise, overlapping speech)? Double that cleanup time. The “4-6 hours to transcribe manually” stat from pre-AI days drops to about 20-40 minutes total with AI assistance – still a massive time saver, just not instant.