Speech Recognition in 2026: Use Cases, Open Source Models and APIs Compared
Updated September 2026.
01Introduction
Speech recognition, also called automatic speech recognition (ASR) or speech-to-text, is technology that converts spoken audio into written text. Developers add it to products in one of two ways: by running an open-source model such as Whisper on their own infrastructure, or by sending audio to a speech-to-text API and getting a transcript back.
Accuracy stopped being the main differentiator a while ago. Most modern models transcribe clear English well. What separates the options now is price per minute, whether you need live or recorded transcription, extras like speaker labels and timestamps, language coverage, and how much infrastructure you are willing to operate. This guide covers what speech recognition is used for, the open-source models worth knowing, the main APIs and what they cost, and how to choose.
02What speech recognition can do
Modern speech recognition covers more than dictation. The capabilities you will see across providers:
- Transcription of recorded audio. Upload a file, get the text. The most common and cheapest mode.
- Real-time transcription. Stream audio and receive text as people speak, for live captions, voice agents and call assistance.
- Speaker diarization. Label who said what in a conversation with several speakers.
- Timestamps. Word or segment timings, needed for subtitles and for jumping to a moment in a recording.
- Language identification. Detect which language is being spoken before or during transcription.
- Speech translation. Transcribe audio in one language and return text in another. Not every API offers all of these, and extras are often priced separately, so check the features you need against each provider's documentation before comparing prices.
03Speech recognition use cases
Meetings and calls. Recording and transcribing meetings, sales calls and support calls is now the largest everyday use. Transcripts feed summaries, CRM notes, quality reviews and searchable archives.
Customer support. Contact centers transcribe calls and voice messages so they can attach them to tickets, search past conversations and spot recurring problems. Real-time transcription also powers agent assist tools that suggest answers during a call.
Healthcare. Clinicians dictate notes instead of typing them, and ambient documentation tools listen to the consultation and draft the record. This saves time but also raises the bar for accuracy on medical terms and for data protection.
Media and content. Podcasts, videos and webinars are transcribed for captions, subtitles, show notes and search. Captions are also an accessibility requirement for many publishers.
Voice interfaces and AI agents. Voice search in shopping apps, smart home control and voice agents that answer phone calls all start with speech recognition. For agents, low latency matters as much as accuracy, because every delay is audible to the caller.
Accessibility. Live captions help people who are deaf or hard of hearing follow conversations, and voice input helps people who cannot easily type.
04Open-source speech recognition models
Self-hosting gives you full control over data and no per-minute fees, in exchange for running and scaling the infrastructure yourself.
Whisper from OpenAI is the default open-source choice. It handles many languages, can translate speech into English, and has a large ecosystem of faster ports and wrappers, such as whisper.cpp for running on CPUs and edge devices. The trade-off is operational: good throughput needs GPUs, and long files, batching and real-time streaming are yours to build.
DeepSpeech from Mozilla used to be a popular recommendation. The project is now discontinued and its repository was archived in June 2025, so it is not a good foundation for a new product.
The real cost of self-hosting is rarely the model. It is GPU capacity, queueing for long files, monitoring, upgrades when a better model is released, and handling the edge cases that a managed API already solves. For a small or spiky workload, an API is usually cheaper once you count engineering time.
05Speech-to-text APIs
A speech-to-text API takes an audio file or stream and returns a transcript. You pay per minute of audio and someone else runs the models.
OpenAI
OpenAI offers several transcription models through the same /v1/audio/transcriptions endpoint. Its newest model, gpt-transcribe, is priced at $0.0045 per minute and supports context prompts, keyword hints and language hints. The original whisper-1 model costs $0.006 per minute. For live audio, gpt-realtime-whisper costs $0.017 per minute.
| OpenAI model | Mode | Price per minute | Price per hour |
|---|---|---|---|
| gpt-transcribe | Recorded files | $0.0045 | $0.27 |
| whisper-1 | Recorded files | $0.006 | $0.36 |
| gpt-realtime-whisper | Real-time streaming | $0.017 | $1.02 |
To put that in numbers: 1,000 hours of recorded audio a month costs $270 with gpt-transcribe and $360 with whisper-1. Moving the same volume to real-time streaming would cost $1,020, so only stream when the use case needs text while people are still talking.
Where it lags: file uploads are limited in size, so long recordings need to be split, and features like speaker labels depend on which model you choose.
AWS, Azure and Google Cloud
Amazon Transcribe fits teams already on AWS, with speaker identification, custom vocabularies and direct integration with S3 and other AWS services. Azure AI Speech supports custom speech models trained on your own data, which helps with specialized vocabulary. Google Cloud Speech-to-Text covers a wide range of languages and integrates with the rest of Google Cloud.
Where they lag: pricing and configuration are more complex than single-purpose APIs, and each ties you further into one cloud. They make the most sense when your data already lives there.
Deepgram and AssemblyAI
Deepgram focuses on fast, real-time transcription and voice agents, with streaming built in from the start. AssemblyAI pairs transcription with audio intelligence features such as summarization, sentiment analysis and content moderation.
Where they lag: both are specialist vendors, which means another contract, another API key and another compliance review for a team that only needs occasional transcription.
Prices for these providers change often and depend on model, volume and features, so compare them on their own pricing pages against your actual monthly minutes.
ApyHub
ApyHub is a curated API catalog where speech recognition sits next to the other audio, text and file APIs a product needs, under one subscription and one API key.
The Convert Speech to Text API transcribes a WAV file, or a public WAV URL, in the language you specify:
curl -X POST "https://api.eu.apyhub.com/apyhub/transcribe-speech-to-text/file" \
-H "apy-token: $APY_TOKEN" \
-F "file=@/path/to/audio.wav" \
-F "language=en-US"{
"data": "Hello, welcome to the meeting."
}Around it you can chain the rest of an audio workflow without another vendor: the MP3 to WAV Converter API to prepare the file, the AI Audio Language Detection API to find out which language is spoken, the AI Video Transcriber API to route audio and video through Azure or Google transcription, and the Convert Text to Speech API to turn text back into audio. Every endpoint is also available through ApyHub MCP, so AI agents can discover, evaluate and call these APIs directly without a hand-written wrapper or tool definition.
Where it lags: the Convert Speech to Text API accepts WAV audio only and returns plain text, without speaker labels, timestamps or confidence scores. The AI Video Transcriber API requires your own Azure or Google credentials. For long recordings that need diarization or live streaming, a specialist provider is the better fit today.
Try the Speech to Text API free →
06How to choose
| If you need | Start with |
|---|---|
| Full control of data, no per-minute fees, and a team that can run GPUs | Self-hosted Whisper |
| The lowest price for recorded files at scale | OpenAI gpt-transcribe, then compare specialists at your volume |
| Live transcription for voice agents or calls | A real-time API such as OpenAI's realtime models or Deepgram |
| Transcription plus summaries, sentiment and moderation | AssemblyAI |
| Everything inside the cloud you already use | Amazon Transcribe, Azure AI Speech or Google Cloud Speech-to-Text |
| Occasional transcription alongside other file, text and audio APIs | ApyHub |
Three questions settle most decisions. How many audio minutes will you process each month? Do you need text while people are still talking, or after the recording ends? And which extras, such as speaker labels, timestamps or specific languages, are requirements rather than nice-to-haves? Test your two best candidates on 20 minutes of your own audio, because accuracy on your accents, vocabulary and recording quality matters more than any benchmark.
07Conclusion
Speech recognition is now accurate and cheap enough to add to almost any product. Recorded transcription costs well under a cent per minute from the leading APIs, and open-source models like Whisper make self-hosting a real option for teams with the infrastructure to run it.
The choice comes down to volume, latency and the extras you need. For recorded audio at scale, compare per-minute prices on your own files. For live use cases, budget for streaming rates. And if speech is one of many capabilities your product needs, consolidating it with the rest of your APIs saves you a vendor, a contract and a compliance review.
08FAQ
What is speech recognition? Speech recognition, or automatic speech recognition (ASR), is technology that converts spoken audio into written text.
What is the difference between speech recognition and voice recognition? Speech recognition identifies what was said. Voice recognition identifies who is speaking, for example to verify a user by their voice.
How much does the OpenAI Whisper API cost? The whisper-1 model costs $0.006 per minute of audio, or $0.36 per hour. OpenAI's newer gpt-transcribe model costs $0.0045 per minute.
What is the best open-source speech recognition model? Whisper is the most widely used open-source model in 2026. Mozilla's DeepSpeech, once a common choice, was discontinued and archived in 2025.
Should I use an API or self-host a model? Use an API if your volume is small or unpredictable, or if you do not have a team to run GPU infrastructure. Self-host if you process very large volumes, need audio to stay on your own servers, or want to avoid per-minute fees.
What is the difference between batch and real-time transcription? Batch transcription processes a recording after it ends and is the cheaper option. Real-time transcription returns text while people are speaking, which voice agents and live captions need, and it usually costs more per minute.
What is speaker diarization? Diarization labels which speaker said each part of a transcript, which is essential for meetings, interviews and call recordings.
Which audio formats do speech-to-text APIs accept? Most accept common formats such as WAV, MP3 and MP4, but some, including ApyHub's Convert Speech to Text API, accept WAV only. Convert other formats first with a tool like the MP3 to WAV Converter API.
09About ApyHub
ApyHub is a curated API catalog and trusted operational layer. It gives developers and AI agents access to 450+ services and 1,500+ endpoints through a single subscription, with new APIs and providers onboarded continuously. Usage is billed in atoms, a unit that reflects the actual work of each call and is pooled across the entire catalog.
Every endpoint ships with machine-readable certification covering data residency, retention, sub-processors and alignment with GDPR, SOC 2 and ISO 27001, and is MCP-ready by default so agents can discover and call it without custom glue code.
ApyHub is headquartered in Amsterdam, with offices in the Netherlands, Greece and India. More than 65,000 developer workspaces use the catalog every month. There is a free tier and no card is required to start.
If you build APIs, you can publish yours to the catalog at apyhub.com/become-a-provider.