Direct answer: public speech-to-text prices in 2026 are roughly: OpenAI gpt-4o-mini-transcribe at USD 0.003/minute, gpt-4o-transcribe/Whisper at USD 0.006/minute, and real-time transcription at USD 0.017/minute; AWS Transcribe batch at USD 0.006/minute, streaming at USD 0.01/minute, and Call Analytics from USD 0.03/minute; Google Speech-to-Text V2 standard at USD 0.016/minute (first 500k minutes), dynamic batch at USD 0.003/minute. At 100,000 minutes/month, the API is roughly USD 300–1,600, but human correction, stereo billing, PII redaction, storage, and integration often cost more. This article covers transcription only, not the full customer-service seat/robot package.
This guide is for operations, customer service, IT/data, and procurement teams building meeting transcription, call QA, sales-call analytics, inspection voice notes, or voice agents. Data cut-off: 31 October 2026. Prices come from official OpenAI, AWS, and Google Cloud pages. Azure Speech did not provide a stable public unit price during this check and is excluded from the numeric comparison. All prices exclude tax; USD and RMB are not converted and no ranking is implied. This is public-source research and an editorial model, not a quote, and it does not promise accuracy, savings, or a go-live date.
Batch, streaming, and real-time are three different prices. Google bills each audio channel separately; AWS bills up to two channels by total duration. Human correction, PII redaction, and storage often cost more than the transcription API.
What are the official speech-to-text minute prices in 2026?
Batch is about USD 0.003–0.006/minute, streaming about USD 0.01–0.017/minute, and Call Analytics from about USD 0.03/minute. Separate batch/streaming/real-time first.
There are at least four ways to buy transcription: asynchronous batch files, streaming real time, call analytics (speakers/emotion/summary), and enhanced versions with PII redaction or custom language models. Prices and units differ; a single “per minute” number is not a full comparison.
| Service / mode | Billing unit | Official list price | Note |
|---|---|---|---|
| OpenAI gpt-4o-mini-transcribe | Minute | USD 0.003/minute | Official estimate; per audio minute |
| OpenAI gpt-4o-transcribe / Whisper | Minute | USD 0.006/minute | Whisper and gpt-4o-transcribe estimates match |
| OpenAI gpt-4o-transcribe-diarize | Minute | USD 0.006/minute | Includes speaker diarization |
| OpenAI real-time transcription | Minute | USD 0.017/minute | gpt-live-transcribe / gpt-realtime-whisper |
| AWS Transcribe batch | Minute | USD 0.006/minute | US East (N. Virginia) example; per-second billing, no minimum |
| AWS Transcribe streaming | Minute | USD 0.01/minute | Includes custom vocabulary, diarization, language ID |
| AWS Transcribe Call Analytics | Minute | From USD 0.03/minute (first 250k) | Tiered; PII redaction/generative summary extra |
| Google Speech-to-Text V2 standard | Minute | USD 0.016/minute (first 500k) | 500k–1M 0.01; 1M–2M 0.008; above 2M 0.004 |
| Google dynamic batch | Minute | USD 0.003/minute | Lower-priority batch |
| Google V1 (with data logging) | Minute | USD 0.016/minute after 60 min | First 60 min/month free; without data logging 0.024 |
Official list prices in USD, excluding tax, as of 31 October 2026. OpenAI uses official estimated cost per minute; AWS uses US East (N. Virginia) examples; Google uses V2 standard/dynamic batch/V1 prices. Azure Speech did not provide a stable public unit price and is excluded. Regional, currency, tier, and add-on terms follow the official page and contract.
What do 10k, 100k, and 1M minutes cost per month?
Batch is about USD 30–160 per 10k minutes; 100k is about USD 300–1,600; 1M is about USD 3,000–13,000, with streaming and Call Analytics higher.
| Option | 10,000 minutes | 100,000 minutes | 1,000,000 minutes | Mode |
|---|---|---|---|---|
| OpenAI gpt-4o-mini-transcribe | USD 30 | USD 300 | USD 3,000 | Batch/per minute |
| OpenAI gpt-4o-transcribe / Whisper | USD 60 | USD 600 | USD 6,000 | Batch/per minute |
| OpenAI real-time | USD 170 | USD 1,700 | USD 17,000 | Real-time |
| AWS batch | USD 60 | USD 600 | USD 6,000 | Async |
| AWS streaming | USD 100 | USD 1,000 | USD 10,000 | Real-time |
| AWS Call Analytics | USD 300 | USD 3,000 | USD 21,450 | Call analytics/tiered |
| Google V2 standard | USD 160 | USD 1,600 | USD 13,000 | Standard |
| Google dynamic batch | USD 30 | USD 300 | USD 3,000 | Low-priority batch |
Editorial model: minutes × unit price; excludes tax, storage, human correction, PII redaction, summarization, integration, and retries. AWS Call Analytics at 1M minutes is modelled as 250k × 0.03 + 750k × 0.0186 = USD 21,450. Google V2 at 1M is 500k × 0.016 + 500k × 0.01 = USD 13,000. OpenAI uses official estimates; actual bills depend on model and region.
Why can stereo/multichannel double the cost?
Google explicitly bills each channel separately; AWS bills up to two channels by total audio duration. Stereo recordings may be billed as 2× minutes on Google — the most commonly missed difference.
Call recordings, meetings, and interviews often use two channels (agent/customer) or more. Google’s billing note says each audio channel is billed separately: 30 seconds with 4 channels is billed as 120 seconds. AWS says a two-channel conversation is billed only by total audio duration, not separately per channel. OpenAI estimates per audio minute and does not split by channel; confirm multichannel rules in the contract.
| Option | 100k minutes, two channels | Billed as 200k minutes | Note |
|---|---|---|---|
| Google V2 standard | 200,000 minutes | USD 3,200 | Each channel billed separately |
| AWS batch | 100,000 minutes | USD 600 | Up to two channels by total duration |
| OpenAI batch | 100,000 minutes | USD 600 | Per audio minute; confirm multichannel in contract |
Editorial model; excludes tax, correction, storage, and integration. Google’s 200k minutes is still in the first 500k band at USD 0.016/minute; AWS/OpenAI use USD 0.006/minute. Verify multichannel rules with real audio and invoices.
Why is the API fee not the speech project budget?
Human correction, PII redaction, storage, summarization, integration, and exception handling are the main lines; the API only turns audio into text.
| Cost line | Included in API price? | How to meter | Common omission |
|---|---|---|---|
| Transcription API | Yes | Audio minutes × price; channel rules | Stereo, streaming/batch differences |
| Human correction/annotation | No | Audio minutes ÷ review speed × labour rate | Assuming ASR output can be stored directly |
| PII redaction/compliance | Partly | Redacted minutes, audit, permissions | Leaking ID/bank/address data |
| Storage/bandwidth | No | Audio/transcript retention, region, encryption | Keeping raw audio indefinitely |
| Summary/LLM analysis | No | Tokens or per-call | Forgetting summarization cost |
| Integration/write-back | No | CRM/ticket/warehouse APIs, queues, retries | Transcribing without a workflow |
| Maintenance/versioning | No | Model/vocabulary updates, regression | No owner when accents/product names change |
Editorial guidance, not an accounting taxonomy. Follow corporate finance policy, vendor contracts, and actual invoices.
At 100k minutes, how much is human correction?
If 10% needs correction at 2× the audio duration and RMB 150/hour, that is about 333 hours and RMB 50,000 — far above the USD 300–1,600 batch API fee.
| Correction-share assumption | Minutes to correct | Correction hours (2× audio duration) | Labour cost (RMB 150/hour) | Fit |
|---|---|---|---|---|
| 5% | 5,000 | 166.7 hours | RMB 25,000 | Internal meeting notes, low risk |
| 10% | 10,000 | 333.3 hours | RMB 50,000 | Call QA, sales review |
| 100% | 100,000 | 3,333 hours | RMB 500,000 | Medical/legal/compliance verbatim |
100,000 minutes/month. Correction share, 2× audio duration, and RMB 150/hour are editorial assumptions, not an industry standard or quote. API fees are in USD and labour in RMB; they are listed separately and not converted or added. Replace them with your own WER, review speed, and labour cost.
How to choose by scenario
Meeting notes: batch + diarization. Call QA: Call Analytics/PII. Real-time voice agents: streaming. For stereo, confirm channel billing first.
| Scenario | Start with | Why | Do not ignore |
|---|---|---|---|
| Meeting/interview notes | Batch + diarization | Not real-time; diarization separates speakers | Accents, overlapping speech, jargon |
| Call-center QA | Call Analytics/PII redaction | Speakers, sentiment, keywords, redaction | Minute tiers, summary extra, retention |
| Real-time voice agent | Streaming/real-time | Low latency, transcribe while speaking | USD 0.01–0.017/minute, network/interruptions |
| Stereo recordings | Confirm channel billing | Google bills per channel; AWS two channels by total | Do not estimate with a mono price |
| Medical/legal verbatim | High-accuracy model + human | Errors are costly; line-by-line review | Privacy, retention, domain vocabulary |
| Mass offline archive | Dynamic batch/low priority | Lowest price, delay acceptable | Latency, reruns, storage |
Editorial guidance, not a vendor ranking. Start with 1–2 hours of real audio (accents/noise/jargon) and measure WER, diarization, and correction hours before scaling.
What should procurement and acceptance measure?
Measure WER, domain terms, diarization, timestamps, PII redaction, channel rules, and cost per 1,000 minutes — not a vendor demo or unit price alone.
| Acceptance item | Suggested definition | Evidence |
|---|---|---|
| WER/CER | By language/accent/noise/scenario | Labelled audio, scoring script |
| Domain terms | Product/model/material/process term hit rate | Vocabulary, sampling record |
| Diarization | Speaker count/turns/overlap accuracy | Diarization output, manual check |
| Timestamps/segmentation | Sentence/word timestamp usability | Transcript JSON, alignment check |
| PII redaction | ID/bank/phone/address recall and false positives | Redaction sample, audit log |
| Channel rules | Stereo billed at 1× or 2× | Invoice, contract, test audio |
| Cost per 1,000 minutes | (API + correction + storage + integration) ÷ 1,000 minutes | Invoice, hours, project input |
| Data boundary | Region, retention, training use, permissions, DPA | Contract, configuration, audit logs |
There is no universal “good enough” WER; meeting notes and medical verbatim carry different risk. High-risk scenarios need line-by-line review, not a vendor’s average accuracy.
Who should read this — and who should not?
For teams building meeting transcription, call QA, sales-call analytics, or voice agents. Not for a few hours of audio, no correction process, or no compliance plan.
Read if: you work in customer service, sales operations, knowledge management, IT/data, compliance, or procurement; you are comparing OpenAI, AWS Transcribe, and Google Speech-to-Text; you need to compare batch/streaming/real-time, stereo, PII, and human correction. Skip if: you have only a few hours of audio and expect free tiers to solve production; you have no labelled sample or correction process; you treat ASR as a legal/medical verbatim transcript; you mix a full customer-service robot package with a transcription API. Supports: minute-price comparison, 10k/100k/1M-minute models, stereo multiplier, correction TCO, scenario selection, and acceptance metrics.
Lingyi does not resell speech APIs or claim any engine is “most accurate”. If you need transcription integrated with CRM/tickets, a private gateway, permissions, and audit, see LingHarness; for a meeting-notes pilot on public APIs, the cloud vendor is usually faster. To scope speech scenarios, data boundaries, and acceptance, talk to us.
See our separate notes on AI customer-service cost, document OCR cost, and private-deployment TCO; this article only covers speech-to-text APIs.
FAQ
Batch, streaming, or real-time transcription?
Meeting notes and recorded archives use batch, the lowest price; real-time agent assistance and voice agents use streaming/real-time, which costs more but lowers latency; call QA uses Call Analytics/PII. First decide whether the business needs “text while speaking”, then compare price.
Why can stereo double Google’s cost?
Google states each audio channel is billed separately; 100k stereo minutes may be billed as 200k minutes. AWS states a two-channel conversation is billed only by total audio duration, not separately per channel. OpenAI estimates per audio minute; confirm multichannel rules in the contract.
Are OpenAI and AWS batch both USD 0.006/minute?
The public estimates are close, but models, features, regions, free tiers, and add-ons differ. OpenAI has a USD 0.003 mini tier and a USD 0.017 real-time tier; AWS has streaming at USD 0.01 and Call Analytics from USD 0.03. Do not compare one batch price alone.
What does 100,000 minutes cost per month?
Batch is about USD 300–1,600: OpenAI mini 300, OpenAI 4o/Whisper 600, AWS batch 600, Google V2 1,600, Google dynamic batch 300; streaming USD 1,000–1,700; Call Analytics from USD 3,000. Add human correction, PII, storage, and integration; the table is an editorial model, not a quote.
How much human correction is needed?
If 10% needs correction at 2× the audio duration and RMB 150/hour, that is about 333 hours and RMB 50,000; 100% verbatim review is about 3,333 hours and RMB 500,000. The API may be only a few hundred to USD 1,600; correction is the main cost. These are editorial assumptions, not a quote.
How should we accept speech-to-text quality?
Use the same real audio (accents, noise, jargon, stereo) and measure WER/CER, domain-term hits, diarization, timestamps, PII redaction, and cost per 1,000 minutes. Do not rely on a vendor’s average accuracy or demo; medical/legal/compliance content needs line-by-line review.
Method and limitations
This article uses only official pricing pages accessible on 31 October 2026. It does not use private contracts, customer invoices, or vendor internal data. Prices, regions, free tiers, models, and add-ons can change. The article has no actual enterprise agreement, tax, FX, human correction, or storage data, so every monthly cost is an editorial model, not a quote. Azure Speech did not provide a stable public unit price and is excluded. OpenAI multichannel rules and AWS/Google tiers/add-ons follow the contract/invoice. Stereo, correction share, and labour cost are assumptions.
Sources and basis
Data cut-off: 31 October 2026. The following are official pricing/product pages; prices, regions, allowances, and terms follow the original page and contract.
- OpenAI, API Pricing: platform.openai.com/docs/pricing. Per-minute estimates for gpt-4o-transcribe, mini, diarize, Whisper, and real-time transcription.
- AWS, Amazon Transcribe Pricing: aws.amazon.com/transcribe/pricing. Batch 0.006, streaming 0.01, Call Analytics from 0.03, PII/summary add-ons, and two-channel rules.
- Google Cloud, Speech-to-Text pricing: cloud.google.com/speech-to-text/pricing. V2 standard tiers, dynamic batch 0.003, V1 prices, and per-channel billing.