Best Free Speech-to-Text APIs & Open-Source Engines (2026)

‍Quick answer: what's the best free speech-to-text option?

If you want the fastest path to accurate transcription without paying to start, the best free speech-to-text option for most developers is a free-tier API — and AssemblyAI is the strongest of them, giving you free API credit to start (no credit card) plus an open playground to test on your own audio. If you need to run fully offline or self-host with no per-hour cost at all, the best free option is an open-source engine, and OpenAI's Whisper is the most capable one available today. Everything below unpacks that answer: the free-tier APIs, the open-source engines, and how to pick between them.

Here's the honest framing before we start. "Free" means two very different things in speech-to-text. A free-tier API is a hosted service you call over the network — someone else runs the model, you get a limited free allotment, and you pay per hour past it. An open-source engine is model weights and code you run yourself — no per-hour fee, but you own the servers, the GPUs, and the accuracy gap. Which one is right depends entirely on whether you'd rather manage a bill or manage infrastructure.

What is a speech-to-text API?

A speech-to-text API — also called an automatic speech recognition (ASR) API — takes audio in and returns text out over a simple network call. You send a file or an audio stream, the provider runs it through a trained speech-to-text model, and you get back a transcript, usually with timestamps, speaker labels, and confidence scores. You don't host anything, you don't manage GPUs, and you don't retrain models — you make an HTTP request or open a WebSocket and read the response.

That convenience is the whole pitch. The tradeoff is that you're paying per hour of audio and sending data to a third party, which is exactly why some teams reach for open-source engines instead. Both are legitimate. The rest of this guide covers both.

It helps to hold the two categories side by side before diving in, because the decision usually comes down to a handful of dimensions:

Keep that table in mind as you read. Almost every choice below is a point on this spectrum between "manage a bill" and "manage infrastructure."

Free speech-to-text APIs and Voice AI models

These are hosted services with a genuine free tier. The free allotments are small — they're for evaluation, not production — but they let you test real accuracy on real audio before committing.

Free tier and limits at a glance

AssemblyAI

AssemblyAI is a Voice AI platform built around one job done extremely well: turning audio into accurate, structured text, and giving developers the pieces to build on top of it.

The current flagship for recorded audio is Universal-3.5 Pro (model ID universal-3-5-pro) at $0.21/hr. It does native code-switching across 18 languages — meaning it handles a speaker who slips between, say, English and Spanish mid-sentence without breaking — and ships the most accurate diarization we've built, optimized for cpWER, plus contextual prompting to boost domain terms.

For live audio there's Universal-3.5 Pro Realtime at $0.45/hr base, the streaming flagship and the foundation of our Voice Agent API. It uses neural end-of-turn detection that reads tonality and pacing — roughly 300ms — instead of just waiting for silence, which is the difference between a voice agent that feels natural and one that interrupts you.

If you're cost-sensitive or need maximum language coverage, Universal-2 at $0.15/hr is the value tier, covering 99+ languages. It's also a great fallback: you can pass an ordered list, ["universal-3-5-pro", "universal-2"], to default to the flagship and fall back gracefully. (The previous-generation Universal-3 Pro remains available as a pinnable snapshot if you need version stability, but Universal-3.5 Pro is where new projects should start.)

On accuracy, the numbers hold up under scrutiny. On a normalized code-switching benchmark across five language pairs, Universal-3.5 Pro posts a 7.69 average word error rate — ahead of ElevenLabs Scribe v2 (8.77), the prior Universal-3 Pro (9.07), Deepgram Nova-3 (12.22), and OpenAI's GPT-4o Transcribe (44.58). That's a real, measurable edge over the open-source options later in this guide, Whisper included. Full methodology and more results live on the benchmarks page, and if you want to run your own test, here's how accurate speech-to-text really is.

Accuracy is only half the platform. On top of the transcript, AssemblyAI's Speech Understanding models add the structured layer most applications actually need — speaker diarization, sentiment, entity and topic detection, and PII redaction — and the LLM Gateway gives you one endpoint to run summaries, chapters, or custom prompts against any leading model (GPT, Claude, Gemini) without stitching in a separate LLM vendor. That's the difference between a raw transcript and something you can build a product on.

One thing none of the free or open-source alternatives offer: a native voice-agent path. The Voice Agent API replaces a whole STT + LLM + TTS stack with a single WebSocket at a flat $4.50/hr — one meter instead of three. If you're heading toward voice agents, that consolidation is worth knowing about before you wire together three separate vendors.

• Free to start: free API credit at signup, no credit card, plus an open playground.
• Pricing past free: $0.15/hr (Universal-2) to $0.21/hr (Universal-3.5 Pro) async; $0.45/hr streaming; billed per second, no minimums, unlimited concurrency.
• Pros: best-in-class accuracy, native code-switching, real-time streaming, speaker diarization, a native voice-agent path, clean SDKs (pip install assemblyai / npm install assemblyai).
• Cons: hosted only — there's no self-hosted, run-it-on-your-own-hardware option.

Google Speech-to-Text

Google's Cloud Speech-to-Text supports 125+ languages and offers 60 free minutes per month, plus cloud credits for new Google Cloud accounts. If your infrastructure already lives in Google Cloud, it's a reasonable default.

• Free to start: 60 minutes/month, ongoing, plus new-account cloud credits.
• Pros: broad language coverage, tight Google Cloud integration.
• Cons: to transcribe a file it generally has to already live in a Google Cloud Storage bucket, setup is complex, and accuracy trails the leaders on hard audio.

AWS Transcribe

Amazon Transcribe gives you 60 free minutes per month for your first 12 months and includes a specialized Transcribe Medical API. Like Google, it pulls you toward its own ecosystem.

• Free to start: 60 minutes/month, first year only.
• Pros: deep AWS integration, a medical-specific option.
• Cons: audio generally has to sit in an Amazon S3 bucket first, initial setup is fiddly, and accuracy lags the top providers.

How to evaluate speech-to-text solutions

Before you commit to anything — free tier or open source — test it against these five dimensions on your own audio, not a vendor's demo clip.

Accuracy testing for your use case

Word error rate (WER) is the standard metric, but the headline number hides a lot. What matters is accuracy on your audio: your accents, your domain vocabulary, your background noise, your code-switching. A model that scores well on clean read speech can fall apart on a noisy contact-center call.

Use real numbers as your yardstick. On code-switching, Universal-3.5 Pro's 7.69 average WER beats ElevenLabs Scribe v2's 8.77 and Deepgram Nova-3's 12.22 — but the only test that counts is the one you run on a representative sample of your own data. Here's how to evaluate speech recognition models rigorously, including why raw WER can mislead.

Accuracy stat callout — async code-switching WER (lower is better): Universal-3.5 Pro 7.69 · ElevenLabs Scribe v2 8.77 · Universal-3 Pro 9.07 · Deepgram Nova-3 12.22 · OpenAI GPT-4o Transcribe 44.58.

A practical way to run this: assemble 30–60 minutes of audio that looks like your production traffic, hand-correct it into a "truth" transcript, then score each candidate's output against it. Pay attention to entity errors specifically — a name, a phone number, or a dollar amount transcribed wrong costs far more than a dropped filler word, and headline WER treats them the same. That distinction is exactly why we publish separate entity-error and diarization numbers on the benchmarks page.

Feature completeness

Transcription is table stakes. Do you also need speaker diarization, real-time streaming, timestamps, confidence scores, PII redaction, multilingual support, or summarization? A hosted API bundles these; an open-source engine usually means building them yourself. Count the features you actually need before comparing rates.

Be specific about the hard ones. Real-time streaming and natural turn-taking are genuinely difficult to build well — a naive silence-timer approach produces an agent that talks over people. Code-switching (a speaker moving between languages mid-sentence) breaks most models entirely. Diarization that stays accurate on overlapping speech is its own research problem. If your use case touches any of these, weight them heavily, because they're exactly where free tiers and open-source engines tend to fall short of a purpose-built hosted model.

Developer experience and documentation

A clean SDK, honest docs, and predictable errors save more money than a lower rate card. First integrations commonly run 40–80 hours — a provider that halves that has paid for a lot of per-hour rate difference. Test the docs during your evaluation; they don't get better after you commit.

Scalability and reliability

Free tiers cap concurrency and throughput. Ask what production looks like: rate limits, concurrency ceilings, uptime SLAs, regional endpoints. AssemblyAI, for instance, offers pay-as-you-go with unlimited concurrency and no minimums — but always confirm the scaling story before you build on a free tier you'll outgrow in a week.

Total cost of ownership

The per-hour rate is one line item. The full picture includes rework from transcription errors, engineering and maintenance time, storage and egress, and — for self-hosted models — GPUs and MLOps headcount. We break the whole calculation down in our guide to speech-to-text pricing and when to switch providers. The short version: the cheapest rate card is rarely the cheapest total.

Here's how the two paths tend to compare across those five dimensions once you're past a demo and thinking about production:

Open-source speech transcription engines

These have no per-hour fee — you run them on your own hardware. That's genuinely free of licensing cost, but you take on the servers, the scaling, the maintenance, and the accuracy gap. For teams with ML expertise and a data-privacy or offline requirement, that's a fair trade. For most product teams shipping on a deadline, a hosted API wins on total cost.

One caveat worth stating plainly: "open source" and "actively maintained" are not the same thing. Several of the engines below were influential in their day but have been archived or effectively abandoned — building new work on them means starting on a foundation that will never improve. We've flagged the status of each one honestly so you don't inherit a dead dependency.

Whisper (OpenAI)

Whisper is the most capable open-source ASR model available today. It's released under the permissive MIT license — free for commercial use — supports 90+ languages, and ships in sizes from tiny to large-v3 so you can trade accuracy for speed. There's also a hosted Whisper API at $0.006/minute (about $0.36/hr) if you don't want to run it yourself.

• License: MIT (commercial use allowed).
• Pros: strong multilingual coverage, open weights, large community, faster variants like faster-whisper and Distil-Whisper.
• Cons: accuracy has largely plateaued across recent releases, it's comparatively weak on code-switching and diarization, and self-hosting carries real GPU and maintenance overhead.

Kaldi

Kaldi is the veteran research toolkit — Apache 2.0 licensed, production-tested, with a deep and active community. It's powerful and highly customizable, and it can train strong domain-specific models.

• Pros: excellent accuracy when tuned, full control, mature research ecosystem.
• Cons: steep learning curve, command-line and recipe-driven, and expensive to run and maintain at scale.

SpeechBrain

SpeechBrain is a modern PyTorch-based toolkit tightly integrated with Hugging Face, which makes pulling in pre-trained models straightforward.

• Pros: flexible, actively updated, easy access to pre-trained models through Hugging Face.
• Cons: real customization takes ML expertise, and documentation can be sparse in places.

DeepSpeech (Mozilla)

DeepSpeech was Mozilla's open-source engine, designed to run everywhere from a GPU down to a Raspberry Pi 4. Worth being accurate about its status: Mozilla archived the project in late 2022, and it is no longer maintained. You can still use it, but there are no updates, no official support, and no accuracy improvements coming.

• Pros: easy to customize, can train custom models, runs on modest hardware.
• Cons: archived and unmaintained since 2022 — treat it as legacy, not a foundation for new work.

Coqui STT

Coqui STT grew out of the DeepSpeech lineage, supports 20+ languages, and returns confidence scores. Like DeepSpeech, it's no longer actively maintained, so its models won't improve going forward.

• Pros: confidence scores, pre-trained models, an established community.
• Cons: effectively unmaintained; no ongoing model improvements.

Flashlight ASR (formerly Wav2Letter)

Flashlight ASR — the project once known as Wav2Letter, from Facebook AI Research — is a fast C++ engine built on the ArrayFire tensor library.

• Pros: highly customizable, fast processing.
• Cons: complex setup, no ready-made pre-trained model library, and you'll need to source and manage training data yourself.

Understanding free tier limitations and scaling considerations

Common free tier restrictions

Free tiers throttle you on purpose: capped monthly minutes, limited concurrency, sometimes no access to premium models or features, and occasionally watermarked or delayed output. That's fine for evaluation — it's a problem the moment you go to production.

Planning for scale

Model your real monthly volume early. A hobby project transcribing a few hours a month lives comfortably in a free tier or on a self-hosted Whisper instance. A product doing thousands of hours needs pay-as-you-go pricing with real concurrency — and at that scale, the accuracy and maintenance differences between options dwarf the per-hour rate.

Roughly, here's where each type of project lands:

Cost modeling beyond free tiers

When you outgrow free, run the full total-cost-of-ownership math: API rate, plus error-driven rework, plus engineering and maintenance, plus (for self-hosting) GPUs and MLOps. A more accurate model at a higher rate is frequently the cheapest total once rework is counted. Our pricing guide walks through the calculation.

Migration strategies

Whatever you pick first, keep it swappable. Abstract the transcription call behind your own interface so switching providers later is a config change, not a rewrite — free tiers and unmaintained open-source projects are exactly the things you'll want to move off of eventually.

Which free speech-to-text API, Voice AI model, or open-source engine is right for your project?

Cutting through it:

• You want accuracy and speed with the least effort: a free-tier API. Start with AssemblyAI's free credit, test on your own audio in the playground, and scale on pay-as-you-go pricing when you're ready.
• You must run fully offline or self-host for privacy: Whisper. It's the most capable open-source engine, MIT-licensed and commercial-friendly — just budget for GPUs and maintenance.
• You're already deep in a cloud ecosystem: Google or AWS can make sense for the integration convenience, accepting the accuracy and bucket-storage tradeoffs.
• You're building a voice agent: skip stitching three vendors together and look at the Voice Agent API — one WebSocket, flat $4.50/hr, all-in.
• You need a research toolkit to train custom models: Kaldi or SpeechBrain, if you have the ML expertise to run them.

Avoid building anything new on DeepSpeech or Coqui — both are unmaintained, and you'd be starting on a foundation that isn't being improved.

Getting started with speech-to-text integration

Quick start with an API

The fastest way to see real accuracy is to transcribe your own audio. With AssemblyAI it's a few lines:

# pip install assemblyai
import assemblyai as aai
import os

aai.settings.api_key = os.environ["ASSEMBLYAI_API_KEY"]

config = aai.TranscriptionConfig(
    speech_models=["universal-3-5-pro", "universal-2"],  # latest first, stable fallback
    speaker_labels=True,
)
transcript = aai.Transcriber(config=config).transcribe("https://assembly.ai/wildfires.mp3")

if transcript.status == aai.TranscriptStatus.error:
    raise RuntimeError(transcript.error)
print(transcript.text)

Test it on your own audio in minutes. Get your free API key — free credit to start, no credit card — or open the playground to try it with no code.

Integration best practices

Abstract the provider behind your own interface, handle errors and retries from day one, and use a fallback model list so a single hiccup doesn't take down transcription. If you're building in Python, our guide to Python speech recognition covers the libraries and patterns in depth.

Moving to production

Before you ship: confirm concurrency limits and SLAs, budget for volume past the free tier, add monitoring on latency and error rates, and lock in your data-privacy and compliance requirements. If you're multilingual, our overview of multilingual speech-to-text is a good next read.

Ready to build? Try AssemblyAI free or explore the full speech-to-text API.

Frequently asked questions about free speech-to-text APIs

What is the best free speech-to-text option for most developers in 2026?

For most developers, a free-tier API is the best free option because it delivers high accuracy with no infrastructure to manage. AssemblyAI offers free API credit to start (no credit card) plus an open playground; if you need to run fully offline, Whisper is the best open-source engine.

Is there a truly free speech-to-text API, or do they all charge eventually?

Free-tier APIs like AssemblyAI, Google, and AWS give you a limited free allotment for evaluation, then charge per hour past it. For unlimited use with no per-hour fee, you'd self-host an open-source engine like Whisper — free of licensing cost, but you pay for the hardware and maintenance.

What's the best open-source speech-to-text engine right now, and how does it compare to a paid API?

Whisper is the most capable open-source engine — MIT-licensed, 90+ languages, commercially free. But its accuracy has plateaued and it's weak on code-switching and diarization, where a hosted model like Universal-3.5 Pro (7.69 average code-switching WER) pulls ahead.

Is Whisper free for commercial use, and are there any licensing catches?

Yes. Whisper is released under the permissive MIT license, so you can use it commercially, modify it, and self-host it without a licensing fee. Your only real costs are the compute to run it and the engineering to maintain it.

Which free or low-cost speech-to-text API is the most accurate for hard audio?

On a normalized code-switching benchmark, Universal-3.5 Pro leads at 7.69 average WER versus ElevenLabs Scribe v2 (8.77), Deepgram Nova-3 (12.22), and OpenAI GPT-4o Transcribe (44.58). Always confirm by testing candidates on a representative sample of your own audio.

Do free speech-to-text APIs support real-time streaming, or only recorded files?

Many support both. AssemblyAI offers real-time streaming with Universal-3.5 Pro Realtime, which uses neural end-of-turn detection for natural turn-taking — ideal for voice agents. Google and AWS also offer streaming, though free-tier concurrency and throughput are limited.