# Best Speech-to-Text API for Indie Developers (2026)

> Speech-to-text APIs compared for indie devs: batch and streaming $/hr, free credits and billing minimums for OpenAI, Deepgram, AssemblyAI, Groq, Soniox and more.

- Canonical: https://zeroraised.com/blog/best-speech-to-text-api-for-indie-developers
- Published: 2026-11-13T07:30:00Z
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Building
- Tags: ai, api, developer-tools, comparison

**Short answer:** For recorded audio (meeting notes, podcasts, uploads), start with Soniox at $0.10 per hour or AssemblyAI Universal-2 at $0.15 per hour, both with speaker labels for a few cents or less. For live audio and voice agents, start with Deepgram, whose $200 free credit covers a lot of prototyping, or AssemblyAI's streaming tier. If raw price is all that matters and you do not need speaker labels, Groq-hosted Whisper turbo is $0.04 per hour.

Every price below comes from the vendor's own pricing page or docs, checked on 13 Nov 2026. The recommendations are ours and are marked as such.

## TL;DR

- **Cheapest batch:** Groq `whisper-large-v3-turbo` at $0.04 per hour, 10 second minimum per request ([Groq docs](https://console.groq.com/docs/speech-to-text)).
- **Cheapest with speaker labels:** Soniox async at $0.10 per hour, diarization included ([Soniox pricing](https://soniox.com/pricing)).
- **Best free credit:** Deepgram, $200 with no card and no expiry ([Deepgram pricing](https://deepgram.com/pricing)).
- **One vendor for everything:** OpenAI, from $0.003 per minute for `gpt-4o-mini-transcribe` ([OpenAI pricing](https://developers.openai.com/api/docs/pricing)).
- **Watch the billing rules:** minimum billed lengths and session-time billing can cost more than the headline rate on short clips or idle streams.

**Disclosure:** ZeroRaised is a launch site for bootstrapped products. We have no commercial relationship with any vendor below and earn nothing if you sign up. Prices change often, so re-check the linked page before you commit.

## What indie developers actually need from a speech-to-text API

Most indie use cases fall into two buckets, and they want different things.

**Batch (pre-recorded) audio:** meeting notes, podcast transcripts, video captions, voice memos, customer call reviews. You upload a file and get a transcript back. Price per hour, speaker diarization (who said what) and language coverage matter most. Latency barely matters.

**Streaming (real-time) audio:** dictation, live captions, voice input in an app, and voice agents. Audio goes over a WebSocket and words come back as the person speaks. Latency, end-of-turn detection and how the vendor bills open connections matter most.

If you are building the talking half of a voice app, see our [best text-to-speech API guide](https://zeroraised.com/blog/best-text-to-speech-api-for-indie-developers). If you are building a full voice agent, our guide on [how to build an AI voice agent](https://zeroraised.com/blog/how-to-build-an-ai-voice-agent) covers the whole pipeline. This post sticks to turning speech into text.

## Speech-to-text API pricing compared

Prices are converted to dollars per hour of audio where the vendor quotes per minute (per minute × 60). "Not stated" means we could not find it on the official pricing page today, not that it does not exist.

| API (model) | Batch $/hr | Streaming $/hr | Free tier or credit | Billing increment |
|---|---|---|---|---|
| Groq (whisper-large-v3-turbo) | $0.04 | No streaming listed | Free tier (25 MB file limit) | 10 second minimum per request |
| Soniox | $0.10 | $0.12 | Not stated | Token-based, about 30,000 audio tokens per hour |
| AssemblyAI (Universal-2 / Universal-Streaming) | $0.15 (Universal-3.5 Pro $0.21) | $0.15 (Universal-3.6 Pro Realtime $0.45) | $50, no card | Streaming billed on session duration |
| Gladia (Solaria-3) | from €0.16 | from €0.68 | One-time €50 credit, valid 1 year | Not stated |
| OpenAI (gpt-4o-mini-transcribe) | $0.18 ($0.003/min) | $1.02 for gpt-live-transcribe ($0.017/min) | Not stated | Not stated |
| Google Cloud STT V2 | $0.96 standard, $0.18 dynamic batch | $0.96 standard recognition | V1 lists first 60 min/month free | Rounded up to 1 second |
| ElevenLabs Scribe v2/v3 | $0.20 | $0.22 (Scribe v3 Live / v2 Realtime) | Hours bundled into each plan | Not stated |
| Rev AI (Reverb English) | $0.20 | Not covered here | Credit worth 5 hours of Reverb | Per second, 15 second minimum |
| Deepgram (Nova-3 monolingual) | $0.258 ($0.0043/min) | $0.46 ($0.0077/min) | $200, no card, no expiry | Not stated on pricing page |
| Amazon Transcribe (US East example) | $0.36 ($0.006/min) | $0.60 ($0.01/min) | 60 min/month for 12 months | 1 second, no minimum |
| Speechmatics (Melia 1 / Standard / Enhanced) | $0.24 / $0.45 / $0.75 | $0.40 / $0.45 / $0.80 | $100 credit for 12 months | Not stated |
| Azure AI Speech | Price not shown publicly | Price not shown publicly | 5 audio hours/month real-time (F0) | Per second |

Sources: [Groq](https://console.groq.com/docs/speech-to-text), [Soniox](https://soniox.com/pricing), [AssemblyAI](https://www.assemblyai.com/pricing), [Gladia](https://www.gladia.io/pricing), [OpenAI](https://developers.openai.com/api/docs/pricing), [Google Cloud](https://cloud.google.com/speech-to-text/pricing), [ElevenLabs](https://elevenlabs.io/pricing/api), [Rev AI](https://www.rev.ai/pricing), [Deepgram](https://deepgram.com/pricing), [Amazon Transcribe](https://aws.amazon.com/transcribe/pricing/), [Speechmatics](https://www.speechmatics.com/pricing), [Azure](https://azure.microsoft.com/en-us/pricing/details/speech/).

Two notes on the table. Amazon's pricing page renders its tier table with a script, so we used the rates in AWS's own US East worked examples; check your region in the page's dropdown. Azure's public page showed placeholders instead of prices when we checked, so use the Azure pricing calculator for your region.

## The APIs, one by one

### OpenAI: one vendor, sensible prices

OpenAI lists `whisper-1`, `gpt-4o-transcribe` and `gpt-4o-transcribe-diarize` at an estimated $0.006 per minute, `gpt-transcribe` at $0.0045 per minute and `gpt-4o-mini-transcribe` at $0.003 per minute. For live audio, `gpt-live-transcribe` and `gpt-realtime-whisper` are $0.017 per minute ([OpenAI pricing](https://developers.openai.com/api/docs/pricing)).

Our read: if your app already calls OpenAI for the LLM step (summarising a meeting, say), using the same key and bill for transcription is a real simplicity win. `gpt-4o-mini-transcribe` at $0.18 per hour is cheap enough for most indie products. Live transcription is the expensive part of this lineup.

### Deepgram: best free credit, strong streaming

Deepgram's pay-as-you-go rates for Nova-3 monolingual are $0.0043 per minute pre-recorded and $0.0077 per minute streaming; multilingual is $0.0052 and $0.0092. Flux, its streaming model, is $0.0077 per minute for English. Diarization is included for pre-recorded audio and costs $0.0020 per minute extra on streams. New accounts get $200 of credit with no card and no expiry, which Deepgram estimates covers roughly 43,000 minutes of Nova-3. Its Growth plan starts at $4,000 a year of prepaid credit ([Deepgram pricing](https://deepgram.com/pricing)).

Our read: Deepgram is the safest default for a voice product you need to prototype for months without paying.

### AssemblyAI: cheapest mainstream batch with diarization

AssemblyAI charges $0.15 per hour for Universal-2 and $0.21 per hour for Universal-3.5 Pro on pre-recorded audio. Diarization adds $0.02 per hour. Streaming is $0.15 per hour on Universal-Streaming (English or multilingual) and $0.45 per hour base on Universal-3.6 Pro Realtime, plus $0.12 per hour for streaming diarization. You get $50 of credit with no card ([AssemblyAI pricing](https://www.assemblyai.com/pricing)).

The catch is in the billing notes: streaming "is billed on session duration, not audio duration", and "Idle connection time counts." A socket left open for an hour bills an hour. Close connections when the user stops talking.

### Soniox: lowest price that still includes speaker labels

Soniox lists $0.10 per hour async and $0.12 per hour real-time, with diarization, language identification and smart formatting included. It bills in tokens under the hood: about 30,000 input audio tokens per hour of audio ([Soniox pricing](https://soniox.com/pricing)). We could not find a free credit on the pricing page.

### Groq-hosted Whisper: cheapest raw transcription

Groq runs OpenAI's open Whisper models: `whisper-large-v3-turbo` at $0.04 per hour and `whisper-large-v3` at $0.111 per hour. Requests shorter than 10 seconds are billed as 10 seconds, and direct uploads are capped at 25 MB on the free tier and 100 MB on the dev tier ([Groq docs](https://console.groq.com/docs/speech-to-text)).

Our read: great for long files where you only need words. You will need another tool, or your own logic, for speaker labels.

### The others worth a look

- **Google Cloud Speech-to-Text V2:** $0.016 per minute for standard recognition in the first 500,000 minutes, or $0.003 per minute with dynamic batch, which trades speed for price. Requests are rounded up to the nearest second ([Google pricing](https://cloud.google.com/speech-to-text/pricing)). Worth it if you already live on Google Cloud.
- **Amazon Transcribe:** billed in one-second increments with no minimum, 60 free minutes a month for 12 months, and speaker diarization and custom vocabularies included in the standard price ([AWS pricing](https://aws.amazon.com/transcribe/pricing/)).
- **ElevenLabs Scribe:** $0.20 per hour batch and $0.22 per hour real-time, with hours bundled into each plan ([ElevenLabs API pricing](https://elevenlabs.io/pricing/api)). Handy if you already use ElevenLabs for voices.
- **Speechmatics:** Melia 1, its newest multilingual model, lists at $0.24 per hour pre-recorded and $0.40 real-time. Standard is $0.45 per hour for both batch and real-time, Enhanced $0.75 batch and $0.80 real-time, plus $100 of sign-up credit valid 12 months ([Speechmatics pricing](https://www.speechmatics.com/pricing)).
- **Gladia:** async from €0.16 per hour on Solaria-3 and real-time from €0.68 per hour, with a one-time €50 credit. Its pricing table marks diarization as async only ([Gladia pricing](https://www.gladia.io/pricing)).
- **Rev AI:** Reverb English at $0.20 per hour, foreign languages at $0.30 per hour, billed per second with a 15 second minimum ([Rev AI pricing](https://www.rev.ai/pricing)).
- **Azure AI Speech:** 5 free audio hours a month of real-time transcription on the free tier, billed per second ([Azure pricing](https://azure.microsoft.com/en-us/pricing/details/speech/)).

## Worked cost example: 500 hours of recorded audio a month

**Example, our arithmetic on list prices.** Say you run a meeting-notes app and process 500 hours (30,000 minutes) of recorded calls a month, and you want speaker labels.

| API | Calculation | Monthly cost |
|---|---|---|
| Soniox async | 500 × $0.10 (diarization included) | $50 |
| AssemblyAI Universal-2 | 500 × ($0.15 + $0.02 diarization) | $85 |
| Google V2 dynamic batch | 30,000 × $0.003 (diarization not priced here) | $90 |
| OpenAI gpt-4o-mini-transcribe | 30,000 × $0.003 (words only; OpenAI lists diarization as a separate model) | $90 |
| Deepgram Nova-3 pre-recorded | 30,000 × $0.0043 (diarization included) | $129 |
| OpenAI gpt-4o-transcribe-diarize | 30,000 × $0.006 | $180 |
| Speechmatics Standard | 500 × $0.45 | $225 |
| Google V2 standard | 30,000 × $0.016 | $480 |

For comparison, Groq `whisper-large-v3-turbo` would be 500 × $0.04 = $20, but without speaker labels.

The spread from $50 to $480 is the real story. At indie scale, the cheapest sensible option and the expensive one differ by hundreds of dollars a month for the same audio.

### Billing quirk example: short voice clips

**Example, our arithmetic.** A voice-input feature sends 20,000 clips a month, each 5 seconds long. That is 100,000 seconds, or about 27.8 hours of real audio.

- **Groq** bills each clip as 10 seconds: 200,000 seconds, or 55.6 hours × $0.04 = about $2.22. Cheap, but you pay for double the audio.
- **Rev AI** bills each clip as 15 seconds: 300,000 seconds, or 83.3 hours × $0.20 = about $16.67, three times the $5.56 the real audio would cost.
- **Amazon Transcribe** bills per second with no minimum, so you pay for 27.8 hours.

Our read: if your audio is short and bursty, check minimum billed lengths before you check the headline rate. If you stream instead, close connections promptly on session-billed APIs.

## Self-hosting: Whisper, faster-whisper, whisper.cpp and Parakeet

Self-hosting is free in licence terms. OpenAI says Whisper's "code and model weights are released under the MIT License", and the repo lists VRAM needs from about 1 GB for tiny to about 10 GB for large and about 6 GB for turbo ([Whisper on GitHub](https://github.com/openai/whisper)). Community projects such as [faster-whisper](https://github.com/SYSTRAN/faster-whisper) and [whisper.cpp](https://github.com/ggml-org/whisper.cpp) offer alternative Whisper inference engines if the reference code is too slow for you. NVIDIA's [Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) is under CC-BY-4.0, covers 25 European languages and outputs punctuation and timestamps.

Our read: with hosted APIs at $0.04 to $0.15 per hour of audio, a rented GPU rarely pays for itself at indie volumes. Self-host when your audio cannot leave your servers, when you want on-device transcription, or when you already have idle hardware.
## Data, privacy and training

Transcripts of calls and meetings are sensitive. Before you send customer audio to any API, read the vendor's data retention and model-training terms on the day you sign up, and tell your users which processor you use. We cover the main AI vendors' defaults in [do AI APIs train on your data](https://zeroraised.com/blog/do-ai-apis-train-on-your-data). This is not legal advice.

## How to choose in ten minutes

This is our editorial advice based on the prices above.

1. **Recorded audio, need speaker labels, tight budget:** Soniox or AssemblyAI Universal-2.
2. **Recorded audio, words only, lowest cost:** Groq `whisper-large-v3-turbo`.
3. **Live captions, dictation or a voice agent:** Deepgram (prototype on the $200 credit) or AssemblyAI Universal-Streaming. Test latency from your users' region.
4. **Already on OpenAI, Google Cloud or AWS:** use their transcription first; one bill and one set of keys is worth a little extra.
5. **Audio cannot leave your servers:** self-host Whisper or Parakeet.
6. **Many short clips:** prefer per-second billing with no minimum.

Then run 30 minutes of your real audio (accents, background noise, jargon) through your top two. Accuracy on your own audio beats any benchmark table.

## Frequently asked questions

### What is the cheapest speech-to-text API in 2026?

On the official pages we checked on 13 Nov 2026, Groq-hosted whisper-large-v3-turbo is the cheapest at $0.04 per hour of audio, with a 10 second minimum per request. Soniox async is next at $0.10 per hour and includes speaker diarization. Self-hosting Whisper or NVIDIA Parakeet costs nothing in licence fees but you pay for the GPU and your time.

### Which speech-to-text API has the best free tier?

Deepgram has the largest free credit we found: $200 on pay as you go, no credit card and no expiry, which Deepgram estimates at roughly 43,000 minutes of Nova-3. Speechmatics gives $100 of credit for 12 months, AssemblyAI $50 with no card, Gladia a one-time 50 euro credit, and Rev AI credit worth 5 hours of Reverb. Amazon Transcribe gives 60 minutes a month for 12 months and Azure 5 audio hours a month of real-time transcription.

### Is streaming transcription more expensive than batch?

Usually, yes. Deepgram Nova-3 monolingual is $0.0043 per minute pre-recorded against $0.0077 per minute streaming, and OpenAI lists $0.003 to $0.006 per minute for file transcription against $0.017 per minute for gpt-live-transcribe. AssemblyAI and Soniox are exceptions with cheap streaming tiers, but AssemblyAI bills streaming on session duration, so idle open connections cost money.

### Should an indie developer self-host Whisper instead of using an API?

Only if you already have a GPU doing little else, or you have strict data rules. Whisper's code and weights are MIT licensed and NVIDIA's Parakeet TDT 0.6B v3 is CC-BY-4.0, so licensing is not the obstacle. The obstacle is cost and effort: hosted APIs start at $0.04 to $0.15 per hour of audio, which is hard to beat with a rented GPU at indie volumes.

### Which speech-to-text API is best for a voice agent?

Pick a streaming model built for turn-taking and low latency, such as Deepgram Flux or Nova-3, AssemblyAI Universal-Streaming, or OpenAI's live transcription models, then test latency from your own region. Our [voice agent guide](https://zeroraised.com/blog/how-to-build-an-ai-voice-agent) covers the full pipeline, including the LLM and text-to-speech parts.

## Ship it and launch it

Once your transcription feature works, run through our [product launch checklist for bootstrapped makers](https://zeroraised.com/blog/product-launch-checklist-for-bootstrapped-makers), then [launch it on ZeroRaised](https://zeroraised.com/submit), the launch site for products built without VC. For more guides like this, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter).

## Sources

All last checked 13 Nov 2026.

- [OpenAI API pricing](https://developers.openai.com/api/docs/pricing)
- [Deepgram pricing](https://deepgram.com/pricing)
- [AssemblyAI pricing](https://www.assemblyai.com/pricing)
- [Soniox pricing](https://soniox.com/pricing)
- [Groq speech-to-text docs](https://console.groq.com/docs/speech-to-text)
- [Google Cloud Speech-to-Text pricing](https://cloud.google.com/speech-to-text/pricing)
- [Amazon Transcribe pricing](https://aws.amazon.com/transcribe/pricing/)
- [Azure AI Speech pricing](https://azure.microsoft.com/en-us/pricing/details/speech/)
- [ElevenLabs API pricing](https://elevenlabs.io/pricing/api)
- [Speechmatics pricing](https://www.speechmatics.com/pricing)
- [Gladia pricing](https://www.gladia.io/pricing)
- [Rev AI pricing](https://www.rev.ai/pricing)
- [OpenAI Whisper on GitHub](https://github.com/openai/whisper)
- [NVIDIA Parakeet TDT 0.6B v3 model card](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3)
- [faster-whisper on GitHub](https://github.com/SYSTRAN/faster-whisper)
- [whisper.cpp on GitHub](https://github.com/ggml-org/whisper.cpp)

## Frequently asked questions

### What is the cheapest speech-to-text API in 2026?

On the official pages we checked on 13 Nov 2026, Groq-hosted whisper-large-v3-turbo is the cheapest at $0.04 per hour of audio, with a 10 second minimum per request. Soniox async is next at $0.10 per hour and includes speaker diarization. Self-hosting Whisper or NVIDIA Parakeet costs nothing in licence fees but you pay for the GPU and your time.

### Which speech-to-text API has the best free tier?

Deepgram has the largest free credit we found: $200 on pay as you go, no credit card and no expiry, which Deepgram estimates at roughly 43,000 minutes of Nova-3. Speechmatics gives $100 of credit for 12 months, AssemblyAI $50 with no card, Gladia a one-time 50 euro credit, and Rev AI credit worth 5 hours of Reverb. Amazon Transcribe gives 60 minutes a month for 12 months and Azure 5 audio hours a month of real-time transcription.

### Is streaming transcription more expensive than batch?

Usually, yes. Deepgram Nova-3 monolingual is $0.0043 per minute pre-recorded against $0.0077 per minute streaming, and OpenAI lists $0.003 to $0.006 per minute for file transcription against $0.017 per minute for gpt-live-transcribe. AssemblyAI and Soniox are exceptions with cheap streaming tiers, but AssemblyAI bills streaming on session duration, so idle open connections cost money.

### Should an indie developer self-host Whisper instead of using an API?

Only if you already have a GPU doing little else, or you have strict data rules. Whisper's code and weights are MIT licensed and NVIDIA's Parakeet TDT 0.6B v3 is CC-BY-4.0, so licensing is not the obstacle. The obstacle is cost and effort: hosted APIs start at $0.04 to $0.15 per hour of audio, which is hard to beat with a rented GPU at indie volumes.

### Which speech-to-text API is best for a voice agent?

Pick a streaming model built for turn-taking and low latency, such as Deepgram Flux or Nova-3, AssemblyAI Universal-Streaming, or OpenAI's live transcription models, then test latency from your own region. Our voice agent guide covers the full pipeline, including the LLM and text-to-speech parts.