Speech-to-Text API Pricing Models Explained (2026)

Illustration of speech-to-text API pricing, showing an audio waveform branching into batch and real-time transcription with cost indicators.

Understand speech-to-text API pricing, batch vs real-time rates, billing rules, and cost estimates with Smallest.ai Pulse and 12 provider examples.

Speech-to-text API pricing refers to the cost of converting recorded or live audio into text. Smallest.ai Pulse supports both pre-recorded and real-time transcription, making it a practical example for understanding how processing modes, audio duration, and features affect transcription costs.

The total cost depends on more than the advertised rate. Audio duration, processing mode, billing increments, and usage volume affect the final bill. Features such as speaker diarization and redaction may be included in the base rate or charged separately, depending on the provider.

In this guide, you'll learn how speech-to-text API pricing works, how to calculate transcription costs using Smallest.ai Pulse, and how pricing structures differ across 12 providers.

What Is Speech-to-Text API Pricing?

A speech-to-text (STT) API, also called an automatic speech recognition (ASR) API, takes an audio file or live audio stream and returns a transcript. Developers use STT APIs to build voice agents, meeting notetakers, call analytics, live captions and voice search.

When comparing speech-to-text APIs, consider three main pricing factors:

  1. A base transcription rate, charged per unit of audio, such as per minute or per hour.

  2. A processing mode, such as batch (pre-recorded audio) or real-time (streaming), which can affect the transcription rate.

  3. Optional features, such as speaker labels or redaction, which some providers include and others charge for separately.

Smallest.ai Pulse supports features such as speaker diarization, word-level timestamps, and optional PII and PCI redaction. Redaction is currently supported for English and Hindi audio and must be enabled when required. When estimating transcription costs, developers should check which features are available for their chosen model and processing mode, rather than comparing base rates alone.

Understanding Speech-to-Text API Pricing With Smallest.ai Pulse

Smallest.ai offers three speech-to-text models for different transcription needs. Pulse supports multilingual batch and real-time transcription, while Pulse Pro is designed for high-accuracy English transcription of pre-recorded audio. Pulse 2.0 supports English-only real-time transcription with built-in end-of-turn detection, making it suitable for voice agents and live conversations.

To estimate transcription costs with Smallest.ai, select the model and processing mode, then multiply the billable audio minutes by the applicable per-minute rate. Review any additional service charges and plan limits to estimate total usage costs. This helps developers compare transcription options and plan their monthly API budget.

How Speech-to-Text APIs Are Billed

Nearly every speech-to-text API bills on the duration of audio processed, not on the number of words in the transcript. A silent 10-minute recording and a 10-minute conversation usually cost the same.

Billing units: per minute, per hour and per token

Providers quote prices in one of three units:


Unit

How it works

Providers that use it

Per minute

A fixed rate for each minute of audio

Smallest.ai, Deepgram, Google Cloud, Amazon Transcribe, OpenAI

Per hour

A fixed rate for each hour of audio, pro-rated

AssemblyAI, Speechmatics, Azure, ElevenLabs, Gladia, Rev AI

Per token

Audio and text are counted as tokens, then converted to an estimated hourly rate

Soniox

To compare them, convert everything to the same unit. One hour is 60 minutes, so divide an hourly rate by 60:

$0.24 per hour / 60 = $0.004 per minute
$0.15 per hour / 60 = $0.0025 per minute

Billing increments

The billing increment is the smallest slice of audio a provider charges for. Per-second billing means a 14-second clip costs 14 seconds. Coarser rounding means you pay for time you did not use.

Deepgram, Amazon Transcribe and Azure bill per second. Google rounds each request up to the next whole second. Speechmatics consumes credits down to the second.

Minimum charges

Some providers set a minimum per request. Rev AI, for example, rounds each automated file up to the nearest second with a 15-second minimum, so a 5-second voice command is billed as 15 seconds. Minimums matter most for apps that send many short clips.

Session-based billing for streaming

Some real-time speech-to-text APIs charge based on the total streaming session duration rather than only the audio processed. For example, AssemblyAI bills for the entire time a WebSocket connection remains open, including idle periods. Developers should close unused streaming connections to avoid unnecessary charges.

Batch vs Real-Time Transcription Pricing

Speech-to-text API pricing often differs between batch and real-time transcription. Batch processing handles pre-recorded audio, while real-time transcription processes speech as it happens. The cost difference depends on the provider, selected model, and included features.

Batch (pre-recorded) transcription

Batch transcription processes pre-recorded audio and returns a transcript after processing is complete. Depending on the API, developers can submit audio files through an HTTP request and receive results directly or asynchronously through a webhook. Smallest.ai Pulse supports pre-recorded transcription through its speech-to-text API.

Common use cases include call recordings, podcasts, interviews, meeting archives, and media libraries. Batch transcription is often more affordable than real-time processing because it does not require continuous live transcription. However, the actual cost difference depends on the provider, model, and pricing plan.

Real-time (streaming) transcription

Real-time transcription returns partial and final transcripts while audio is still arriving, usually over a WebSocket connection. Voice agents, live captions, agent-assist tools and voice assistants depend on it.

Real-time transcription often costs more than batch transcription because it processes audio continuously and must return results with low latency. However, the price difference varies by provider and model. Some APIs offer similar base rates for both processing modes. Latency, measured as the time from speech to the first transcript, becomes part of what you pay for.

Which mode does your use case need?

  • Use batch when a transcript can arrive seconds or minutes after the audio ends: call analytics, compliance review, podcast transcripts, searchable archives.

  • Use real-time when the transcript drives what happens next: voice agents, live captions, real-time agent assist, voice-controlled apps.

  • Use both when a product needs live text during a call and a cleaner record afterward. Many teams stream for the live experience and re-run the recording in batch for the final transcript.

What Affects the Cost of a Speech-to-Text API?

Beyond batch versus real-time, seven factors shape what you pay.

1. Model tier

Speech-to-text providers may offer multiple models with different accuracy, speed, language support, and pricing. Higher-accuracy models can cost more, but the final rate also depends on the processing mode and features included.

  • Smallest.ai offers three speech-to-text model options. Pulse supports multilingual batch and real-time transcription. Pulse Pro is designed for high-accuracy English transcription of pre-recorded audio. Pulse 2.0 is built for English real-time transcription and includes end-of-turn detection. Choosing the appropriate model helps developers balance accuracy, latency, features, and processing costs.

  • AssemblyAI charges $0.15/hr for Universal-2 and $0.21/hr for Universal-3.5 Pro on pre-recorded audio.

2. Language coverage

Some models serve one language; others handle dozens and switch between them mid-sentence (code-switching). Language support can affect transcription costs because some APIs use different models or rates for multilingual audio. Smallest.ai Pulse supports multilingual transcription, with language availability depending on whether audio is processed in batch or real time. Developers should verify that their required languages and regions are supported before estimating costs.

3. Features and add-ons

Features beyond plain text are where prices diverge most. Common ones include:

  • Speaker diarization: labels who said what in multi-speaker audio.

  • Word-level timestamps: start and end times for each word, used for captions and search.

  • PII and PCI redaction: masks personal and payment-card data in transcripts.

  • Keyword boosting: improves recognition of names, products and jargon.

  • Audio intelligence: sentiment, summaries, entity detection and topic tagging.

Some transcription APIs include several features within their base rates, while others charge for specific add-ons. Smallest.ai Pulse supports speaker diarization, timestamps, keyword boosting, and optional sensitive-data redaction. Redaction is currently supported for English and Hindi when using Pulse, and is not available through Pulse Pro.

4. Audio channels

Audio channel configuration can affect transcription costs. Some providers charge separately for each channel processed, which means a 10-minute stereo recording may be billed as 20 minutes of audio. Other providers, such as Amazon Transcribe, include up to two channels within the standard transcription rate. Developers should check multichannel billing rules before estimating costs.

5. Volume

Some speech-to-text API providers offer lower rates for higher usage volumes, either through automatic usage tiers or committed pricing plans. For example, Google Cloud Speech-to-Text V2 Standard pricing starts at $0.016 per minute and drops to $0.004 per minute at its highest published usage tier, above 2 million minutes per month. Smallest.ai also offers custom Enterprise pricing for higher-volume workloads. Developers should compare standard rates with any available volume discounts or contractual commitments.

6. Concurrency

Concurrency is how many streams or requests can run at once. Self-serve plans cap it, for example at 100 concurrent speech-to-text streams on Smallest.ai pay as you go and 150 on Deepgram pay as you go. Higher limits typically require a growth or enterprise plan.

7. Compliance and deployment

Healthcare, finance and government workloads often need HIPAA agreements, zero data retention, private cloud or on-premise deployment. These usually sit on enterprise plans or carry a fee. Smallest.ai, for instance, charges $1,000 a month for HIPAA zero data retention on pay as you go and includes it, along with on-premise deployment, on Enterprise.

Common Speech-to-Text Pricing Models

Providers package the same per-minute rates in a few standard ways. Knowing which model you are on tells you how predictable the bill will be.


Pricing model

How it works

Example

Pay as you go

Pay the listed rate for what you use, with no commitment

Smallest.ai Pulse, AssemblyAI, Deepgram pay as you go

Prepaid credits

Buy credits up front, often with a bonus or discount, and draw them down

Gladia top-up bonuses, Speechmatics credit packs

Committed plans

Commit to an annual spend for lower rates and higher limits

Deepgram Growth from $4,000 a year, up to 20% off

Volume tiers

The rate drops automatically past set monthly volumes

Google Cloud V2, Amazon Transcribe redaction

Token-based

Audio and text tokens are metered, then expressed as an hourly estimate

Soniox

Enterprise

Custom rates, SLAs, compliance terms and deployment options

Smallest.ai Enterprise, with on-premise deployment and a 99.99% uptime SLA

Speech-to-Text API Pricing: 12-Provider Reference Table

Speech-to-text API pricing varies by transcription model, processing mode, and included features. The table below compares published prices from 12 providers, with Smallest.ai Pulse as the primary example. Use these rates as a reference when estimating costs, and check the applicable model, billing plan, and add-on charges before making a decision.


Provider

Batch (pre-recorded)

Real-time (streaming)

Speaker diarization

Smallest.ai Pulse

$0.003/min ($0.18/hr); Pulse Pro $0.0035/min

$0.004/min ($0.24/hr)

Included

Deepgram

Nova-3 $0.0043/min ($0.26/hr)

Nova-3 or Flux $0.0077/min ($0.46/hr)

+$0.0020/min

AssemblyAI

Universal-2 $0.15/hr; Universal-3.5 Pro $0.21/hr

Universal-Streaming $0.15/hr; Universal-3.6 Pro Realtime $0.45/hr

+$0.02/hr batch; +$0.12/hr streaming

Speechmatics

Melia 1 $0.24/hr list

Varies by model

Included

Google Cloud

Dynamic batch $0.003/min ($0.18/hr)

V2 standard $0.016/min ($0.96/hr), lower at volume

No separate fee listed

Amazon Transcribe

$0.006/min ($0.36/hr)

$0.01/min ($0.60/hr)

Included

Azure AI Speech

$0.18/hr

$1.00/hr

Included in batch; about $0.30/hr in real-time

OpenAI

gpt-transcribe $0.0045/min

gpt-live-transcribe $0.017/min ($1.02/hr)

Not listed on pricing page

ElevenLabs

Scribe v2 $0.22/hr

Scribe v2 Realtime $0.39/hr

No separate fee listed

Soniox

About $0.10/hr

About $0.12/hr

Included

Gladia

Solaria-3 €0.16/hr; Solaria-1 €0.27/hr

€0.68/hr

Included (async only)

Rev AI

Reverb $0.20/hr (English)

Not listed on pricing page

Not listed on pricing page

A few patterns stand out:

  • Batch rates cluster tightly. Most sit between $0.10 and $0.36 per hour, so batch choices often come down to accuracy and features rather than price.

  • Real-time rates spread widely. They run from about $0.12 to $1.02 per hour, a gap of more than 8x.

  • Feature bundling differs. Smallest.ai, Speechmatics, Soniox and Amazon Transcribe include diarization, while Deepgram, AssemblyAI and Azure real-time charge for it.

How to Calculate Your Speech-to-Text API Cost

To calculate your monthly speech-to-text API cost, estimate the audio processed, select the transcription model, and account for additional features and billing rules.

Cost per processing mode = billable audio minutes × (base rate per minute + applicable add-on rates per minute)

Total monthly cost = batch transcription cost + real-time transcription cost + any additional fixed charges

  1. Estimate monthly audio minutes. Multiply sessions or files per month by their average length. For example, 5,000 calls a month at 20 minutes each is 100,000 minutes.

  2. Split by mode. Separate minutes that need real-time transcription from those that can run in batch, since each has its own rate.

  3. List the features you need. Note diarization, redaction, keyword boosting and any audio intelligence, then look up whether each is included or charged.

  4. Adjust for billing rules. Double the minutes for stereo audio if the provider bills per channel, and account for minimums if you send many short clips.

Example 1: Batch transcription for a meeting notetaker

A notetaker app records 10,000 minutes of meetings a month and needs speaker labels. It does not need live text.


Provider

Base rate

Diarization

All-in per minute

Monthly cost

Smallest.ai Pulse

$0.0030

Included

$0.0030

$30.00

For 10,000 minutes of monthly batch transcription, Smallest.ai Pulse's estimated cost is calculated by multiplying the billable minutes by the applicable batch rate. Speaker diarization is supported by Pulse, so developers can enable speaker labels when needed without assuming a separate diarization fee.

Example 2: Real-time transcription for a voice agent

A customer-support voice agent streams 100,000 minutes a month. Every call needs speaker labels and PII redaction.


Provider

Streaming

Diarization

PII redaction

All-in per minute

Monthly cost

Smallest.ai Pulse

$0.0040

Included

Included

$0.0040

$400

AssemblyAI Universal-Streaming

$0.0025

$0.0020

$0.0020

$0.0065

$650

Deepgram Nova-3

$0.0077

$0.0020

$0.0020

$0.0117

$1,170

Amazon Transcribe

$0.0100

Included

$0.0024

$0.0124

$1,240

This example shows why developers should compare the total transcription cost rather than the advertised base rate. Smallest.ai Pulse supports speaker diarization and PII redaction, making it useful for estimating the cost of real-time voice applications that require these features. The final cost depends on the applicable streaming rate, language support, and enabled features.

AssemblyAI's rates are converted from $0.12 per hour per feature, and Amazon's redaction rate is its first tier. Smallest.ai's redaction currently supports English and Hindi audio.

Example 3: Scaling to 1 million minutes

At 1 million audio minutes per month, small differences in per-minute transcription rates can significantly affect the total cost. For example, a difference of $0.001 per minute adds $1,000 to the monthly bill. Smallest.ai Pulse can be used to estimate costs based on the selected processing mode and required features. At this volume, developers should also review custom Enterprise pricing, concurrency limits, and deployment requirements.

Hidden Costs in Speech-to-Text API Pricing


Hidden cost

Why it happens

Example

Add-on stacking

Some providers charge separately for optional features, increasing the total transcription cost.

AssemblyAI Universal-Streaming: $0.15/hr base + $0.12/hr diarization + $0.12/hr PII text redaction = $0.39/hr.

Multichannel audio

Each channel is billed as separate audio

A 10-minute stereo call billed as 20 minutes on AssemblyAI, Google or Deepgram

Idle streaming sessions

Some streaming is billed on connection time

AssemblyAI bills streaming on WebSocket session length

Minimums and rounding

Short requests are rounded up

Rev AI bills a 5-second clip as 15 seconds

Concurrency upgrades

Self-serve caps force a plan change at scale

Smallest.ai supports 100 concurrent STT streams on pay-as-you-go, with custom Enterprise limits. Other providers also set plan-specific concurrency limits.

Compliance terms

Security, data-retention and regulatory requirements may require additional fees, agreements or Enterprise plans.

Smallest.ai lists HIPAA zero-data retention as a $1,000/month pay-as-you-go add-on and includes it in Enterprise. HIPAA Business Associate Agreements and eligibility must be confirmed separately.

Training opt-ins

Some providers offer discounts for allowing model training, while others enable training by default on certain plans.

Speechmatics offers a 33% model-training discount when enabled. Gladia Starter may use customer audio for training by default, but users can opt out. Gladia Growth and Enterprise exclude training by default.

Currency and region

Prices vary by billing currency and cloud region

Gladia bills in euros; Amazon's rates here are for US East (N. Virginia)

Surrounding infrastructure

Storage, logging and data transfer are billed separately

Google bills Cloud Storage separately from Speech-to-Text

The safest way to avoid surprises is to run a week of real traffic through your chosen API with every needed feature enabled, then compare the invoice to your estimate.

How to Reduce Speech-to-Text API Costs

Most savings come from matching the mode, model and features to what each part of your product actually needs.

  1. Send only live-critical audio to real-time. If a transcript can arrive after a call ends, transcribe it in batch. Where batch transcription is cheaper, moving non-urgent audio workloads from streaming to batch can reduce transcription costs.

  2. Prefer bundled features over add-ons. If you need diarization or redaction on most audio, a rate that already includes them, such as Smallest.ai Pulse, is often cheaper than a lower base rate plus per-feature fees.

  3. Choose the model tier per workload. Reserve premium models for audio where accuracy has a clear business cost. With Smallest.ai, developers can use Pulse for multilingual batch or real-time transcription, Pulse Pro for high-accuracy English batch transcription, and Pulse 2.0 for English real-time applications that benefit from built-in end-of-turn detection.

  4. Trim and prepare audio. Remove long silences from recordings before batch transcription, and send mono audio when you rely on diarization rather than separate channels.

  5. Close idle streams. Use voice activity detection and end-of-utterance settings to finalize turns, and close WebSocket connections when a conversation ends.

  6. Group short clips. If your provider has a per-request minimum, combine very short recordings into one request.

  7. Commit once usage is predictable. When monthly volume is steady, prepaid credits, committed plans or enterprise agreements usually beat pay-as-you-go rates.

  8. Monitor usage. Set budget alerts and review usage dashboards weekly so a misconfigured stream or a retry loop is caught early.

Key Takeaways

  • Speech-to-text APIs bill on audio duration, with rates from about $0.10 to $1.02 per hour.

  • Real-time transcription costs more than batch; the premium ranges from about 1.3x to 5.6x depending on the provider.

  • Add-ons such as diarization and redaction can change which provider is cheapest, so compare all-in rates, not base rates.

  • Billing increments, per-channel billing, session-based streaming and minimums all affect the real bill.

  • A simple formula, minutes x (base rate + add-ons), gives a reliable first estimate before you test with real audio.

Estimate Your Transcription Costs With Smallest.ai Pulse

Understanding speech-to-text API pricing starts with knowing your audio volume, processing mode, and required features. Smallest.ai Pulse gives developers a practical way to explore batch and real-time transcription while evaluating the features their applications need.

To get started, explore Smallest.ai Pulse, review the latest model pricing, or follow the speech-to-text API quickstart to transcribe your first audio file.

Frequently asked questions

How much does Smallest.ai Pulse cost?

Which Smallest.ai speech-to-text model should I choose?

What is the difference between per-minute and per-hour pricing?

Does speaker diarization cost extra?

How do I estimate my monthly speech-to-text cost?

Listen to the article
2:00
Listen to the article
2:00

Summarize with AI

Turn Speech into Real-Time Data

Experience fast latency, strong security, and unlimited speech generation.

Summarize with AI

Turn Speech into Real-Time Data

Experience fast latency, strong security, and unlimited speech generation.