What Is Speech to Text? How Transcription AI Works

What is speech to text? It is software that converts spoken audio into written text. Engineers usually call it automatic speech recognition, or ASR. IBM describes it as letting a program convert human speech into written text. Modern systems, such as OpenAI's Whisper, are trained on large amounts of audio and produce transcripts from recordings or live streams.

Speech recognition is the technical name for the same field. It is different from voice recognition, which identifies who is speaking. This guide explains how the technology works, how batch and streaming differ, what affects accuracy, and what it costs in 2026. All prices are a snapshot taken on 17 October 2026.

What is speech to text and how does it work?

Classic systems split the job into stages. According to Wikipedia's speech recognition article, the audio is digitized and turned into spectral features computed in short frames of about 10 milliseconds. An acoustic model maps those features to speech sounds, and a language model scores which word sequences are likely. IBM's explainer describes the same shape: feature extraction, a decoder that uses acoustic models, a pronunciation dictionary and language models, and then word output.

Newer systems learn these parts together. OpenAI describes its Whisper model as a Transformer sequence-to-sequence model that "allow[s] a single model to replace many stages of a traditional speech-processing pipeline." In practice, a modern transcription request follows roughly these steps:

1. The audio is converted into a digital signal and split into short frames.
2. The model turns those frames into candidate sounds or word pieces.
3. The model chooses the most likely word sequence.
4. The system returns text, and some models add punctuation and capitalization.

A short history from Audrey to Transformers

Speech recognition is older than most people assume. The milestones below come from Wikipedia's article and IBM's history notes.

| Year | Milestone | |---|---| | 1952 | Bell Labs built "Audrey", which recognized digits from a single speaker | | 1962 | IBM's 16-word "Shoebox" machine was demonstrated at the World's Fair | | 1971-1976 | DARPA funded a five-year Speech Understanding Research programme | | 1980s | Hidden Markov models replaced dynamic time warping as the main approach | | 1992 | Sphinx-II won DARPA's evaluation as the first speaker-independent, large-vocabulary, continuous recognizer | | 2009-2010 | Deep neural networks were adopted for acoustic modelling | | 2014 | The first end-to-end system based on Connectionist Temporal Classification (CTC) appeared | | 2017 | Microsoft researchers reported human parity on the Switchboard conversational benchmark | | 2022 | OpenAI published the Whisper paper, trained on 680,000 hours of supervision |

Batch vs streaming speech to text

Google Cloud's documentation lists three recognition methods. The difference matters when you pick a tool or API.

| Method | Best for | Detail | |---|---|---| | Synchronous | Short clips | Audio of 1 minute or less; the request waits for the answer | | Asynchronous | Long recordings | Up to 480 minutes; you poll a long-running operation | | Streaming | Live audio | Bi-directional gRPC that returns interim results |

OpenAI's speech to text guide accepts files up to 25 MB in formats including mp3, mp4, m4a, wav and webm. Its live transcription is a separate model, gpt-live-transcribe. A batch job suits recorded interviews or lectures. A streaming connection suits captions for a live event.

What affects speech to text accuracy?

IBM lists pronunciation, accent, pitch, volume and background noise as factors that affect results. It names word error rate (WER) as the main accuracy metric, alongside speed. For the calculation, see word error rate explained.

Benchmarks can mislead if you read them too quickly. In the 2022 Whisper paper, the authors compared Whisper with human transcription services on 25 recordings. The best computer-assisted human service had an aggregate WER 1.15 percentage points better than Whisper, and the authors concluded that English accuracy is "very close to human-level." That is one dataset from 2022, not a general verdict on 2026 systems.

Vendors also publish accuracy percentages, often between 95 and 99 percent. Those figures often come without a shared test set, so treat them as marketing claims until you run the same audio through each tool.

Speech to text, voice recognition and text to speech

People mix up three related terms.

| Term | What it does | Direction | |---|---|---| | Speech to text | Turns spoken words into written text | Audio to text | | Voice recognition | Identifies an individual speaker by voice | Voice to identity | | Text to speech | Turns written text into spoken audio | Text to audio |

IBM separates speech recognition, which turns spoken words into text, from voice recognition, which identifies a user's voice. Text to speech runs in the opposite direction. OndaLabs text to speech covers that side, and the transcription tool covers speech to text.

What does speech to text cost in 2026?

Prices change often, so treat this table as a snapshot. Each figure is a published list price. The per-hour column multiplies the per-minute price by 60.

| Service | Published price | Per audio hour | |---|---|---| | OpenAI gpt-transcribe | $0.0045 per minute | $0.27 | | OpenAI whisper-1 | $0.006 per minute | $0.36 | | Deepgram Nova-3, monolingual | $0.0043 per minute | $0.26 | | Mistral Voxtral Mini Transcribe V2 | $0.003 per minute | $0.18 | | ElevenLabs Scribe v2 | $0.22 per hour | $0.22 | | AssemblyAI Universal-2 | $0.15 per hour | $0.15 | | Google Cloud STT V2 Standard | $0.016 per minute | $0.96 | | OndaLabs, Starter plan (derived) | 50 credits per minute | $0.159 |

The OndaLabs row is derived from the Starter price of $9 for 170,000 credits and the rate of 50 credits per minute. Credits are shared with OndaLabs' other tools. Google's rate applies to its first 500,000 minutes each month, and higher volume tiers cost less. The services differ in languages, features and plan limits, so price alone should not decide your choice.

OndaLabs also includes speech to text on its Free plan, which has 2,000 credits a month. That is about 40 minutes of transcription if every credit goes to it. Full plan details are on the OndaLabs pricing page.

Speaker labels in transcripts

Meeting and interview transcripts often need to show who said what. AssemblyAI's speaker_labels option splits a transcript by speaker, and each utterance gets an ID such as "A". Deepgram adds a speaker index to every word. OpenAI's gpt-4o-transcribe-diarize model is the option OpenAI's docs recommend for speaker labels, priced at $0.006 per minute.

How to transcribe an audio file with OndaLabs

OndaLabs includes a speech to text tool at /app/speech-to-text. It is part of the Free plan, and transcription costs 50 credits per minute. A 10-minute file uses 500 credits, and a 60-minute file uses 3,000 credits, derived from that rate. This article does not make claims about the tool's accuracy, languages or export formats, so check its output in the app before you rely on it.

For open models you can run yourself, read our guide to Whisper alternatives.

Frequently asked questions

What is speech to text used for?

Speech to text is used wherever spoken audio needs to become searchable or editable text. Common examples include meeting notes, captions for video, dictation and call-centre records. Each use case has different needs for speed, speaker labels and accuracy, so the right tool depends on the job.

Is speech to text the same as speech recognition?

Yes. Speech to text is the everyday name, and automatic speech recognition, or ASR, is the technical term. Voice recognition is a different task, because it identifies who is speaking rather than what they said, according to IBM.

How accurate is speech to text in 2026?

Accuracy depends on the audio, the accent, background noise and the vocabulary. The best published comparisons are dataset-specific, such as the 2022 Whisper result on 25 recordings. Vendor accuracy claims are not comparable with each other, so test each tool on your own files and measure word error rate.

What is the difference between real-time and batch transcription?

Batch transcription processes a recorded file and returns the full transcript when the job finishes. Real-time, or streaming, transcription returns interim results while audio is still arriving. Google's asynchronous method accepts recordings of up to 480 minutes.

How much does speech to text cost?

API prices in the table above range from $0.003 to $0.016 per minute, depending on the provider and plan. Consumer tools usually charge a monthly subscription or a rate per hour. OndaLabs charges 50 credits per minute of transcription, and those credits are shared across its tools.