Transcription costs for 10,000 hours: batch, streaming and useful transcripts
Ten thousand audio hours is 600,000 minutes. At the selected Deepgram Nova-3 monolingual pre-recorded rate, that comes to $2,580 before extras. Google's standard V2 volume tiers and its discounted dynamic batch route land on very different totals, so price the route you'll actually run, not the one with the prettiest headline number.
Normalize the audio workload first
Transcribing 10,000 hours of speech is a big job: that's 600,000 minutes of continuous audio. At this scale, small differences in billing meters, volume discounts, and turnaround latency can swing your budget by thousands of dollars. So before you start comparing vendor pricing cards, it helps to pin down a few basics: total audio duration, primary languages, channel count (mono versus stereo), speaker diarization needs, turnaround deadlines, and whether transcription runs asynchronously in batch or in real time during live streams. A batch archive of recorded interviews processed overnight shouldn't be priced on the same tier as an interactive real-time call assistant.
| Selected route | Published basis | 10,000-hour base estimate |
|---|---|---|
| Deepgram Nova-3 monolingual, pre-recorded PAYG | $0.0043 per minute | $2,580 |
| AssemblyAI Universal-3.5 Pro async | $0.21 per hour | $2,100 |
| OpenAI gpt-transcribe | Listed estimated $0.0045 per minute | $2,700 estimated |
| Google Speech-to-Text V2 standard | First 500k minutes at $0.016, next 100k at $0.01 | $9,000 |
| Google V2 dynamic batch | $0.003 per minute with lower processing urgency | $1,800 |
What to keep in mind: these are entry-level base transcription charges, not a like-for-like comparison of output quality or features. Add-ons such as custom vocabularies, PII redaction, entity detection, and structured summaries will show up on your invoice too. It's also worth checking regional processing endpoints and committed-use volume discounts before you lock in a forecast.
Do not flatten a tiered rate into the wrong total
Tiered pricing is one of those spots where a quick spreadsheet formula can quietly mislead you. Under Google Cloud Speech-to-Text V2 standard rates, the volume discount kicks in after 500,000 monthly minutes. For our 600,000-minute workload, the first 500,000 minutes bill at the standard tier rate of $0.016 ($8,000), and the remaining 100,000 minutes bill at the discounted tier rate of $0.01 ($1,000). That gives you an accurate blended total of $9,000. If you multiply all 600,000 minutes by the initial $0.016 rate, you'll get $9,600 and overshoot by $600 of volume tier savings. On the other end, Google's dynamic batch tier drops the price to $0.003 per minute ($1,800 total). The practical catch is that you give up strict turnaround guarantees, so it isn't a drop-in substitute for mission-critical workflows.
One more billing quirk to watch: silent or failed audio. Google and other cloud providers explicitly bill for successfully processed audio files even when the transcript comes back empty because of low input volume or background noise. A little housekeeping in your ingestion pipeline, dropping corrupted or silent recordings before they reach a billable endpoint, can save you from paying for nothing.
Speaker labels can change both cost and usefulness
In multi-party business calls, knowing who said what is often a must-have. AssemblyAI prices asynchronous speaker diarization as an add-on at $0.02 per audio hour on select models. Across our 10,000-hour workload, that adds a tidy $200 on top of the $2,100 base transcription fee, for a total of $2,300 before any downstream summarization. The nuance is that AssemblyAI's real-time streaming diarization sits on a completely different pricing tier, so don't carry async add-on assumptions over into a live agent architecture.
Deepgram keeps separate rate cards for pre-recorded and live streaming audio, and it often highlights temporary promotional pricing for streaming tiers. To keep your forecast honest, budget with the standard list rates. For OpenAI, treat the published estimate of roughly $0.0045 per minute as a guideline rather than a guarantee, especially where the underlying endpoints bill on tokenized audio frames instead of plain audio duration.
Spend a little time on the difficult audio
Before you commit, run your shortlist against your hardest audio: overlapping speakers, heavy industry jargon, cellular compression artifacts, and thick regional accents. Look at both raw word error rate and how much human editing time it takes to make the transcripts usable. Even one minute of manual correction per audio hour adds up to 10,000 minutes, nearly 167 hours of skilled human labor, across a 10,000-hour archive.
Channel setup deserves a quick look too. Some providers bill dual-channel (stereo) files as two separate audio streams, which can double your effective per-minute cost if you upload stereo recordings you don't need. It's worth checking whether server-side channel splitting or automated silence stripping can trim your billable minutes before you send jobs over the wire.
Keep storage and the finished workflow in the quote
A raw transcription API hands you text strings, not a searchable enterprise platform. A production workflow also needs cloud object storage, video playback synchronization, granular access controls, and retention lifecycle policies. If your real goal is capturing team meetings and sharing action items, take a look at specialized meeting-assistant workflow platforms before you take on the engineering effort of building an internal transcription service.
Our worksheet defaults to Deepgram's flat pre-recorded rate, so you can model your own retry factors and add-on charges. For tiered tariffs like Google Cloud V2, calculate each volume bracket separately so the final number stays accurate.
Calculate a flat-rate transcription quote
Starts with Nova-3 monolingual pre-recorded PAYG. Add charged features to the complete per-minute rate or fixed extras. This simple worksheet does not automatically apply volume tiers, channel multipliers or per-request rounding.
Budget the full live voice-agent workflow →