Best Speech-to-Text APIs for Developers Building Voice Apps (2026)

Deepgram is the API most voice-agent teams reach for first because it was built around streaming rather than retrofitted to it. Nova-3 handles general transcription over WebSocket and REST in 45+ languages, with a multilingual mode that copes with code-switching mid-sentence. Flux is the more interesting model for voice apps: instead of just emitting words, it emits Start of Turn, End of Turn, and Eager End of Turn events, so your agent can start preparing a response before the user has fully stopped talking. That's the part of the latency budget most teams struggle to win back with their own voice-activity detection.

The pricing is easy to model. Nova-3 pre-recorded costs about $0.0043 per minute (roughly $0.26 per hour), and streaming is currently $0.0048 per minute on a promotional rate ($0.0077 regular). Billing is per second with no rounding, and the $200 free credit with no card is enough to stream hundreds of hours during evaluation. Word-level timestamps come back on both Nova-3 and Flux, and diarization is included on pre-recorded audio but costs an extra $0.0020 per minute on streams.

Deepgram suits teams whose product lives or dies on conversational feel: phone agents, in-app voice assistants, and live captioning. If you later need TTS or a full agent loop, the same account covers Aura and Flux TTS and a Voice Agent API, so you can consolidate vendors without rewriting the listening layer.

AssemblyAI is the strongest choice when the transcript is the start of your product rather than the end of it. Its model line-up covers both ends of the cost curve: Universal-2 transcribes 99 languages at $0.15 per hour, Universal-3.5 Pro is its most accurate async model at $0.21 per hour across 18 languages with native code-switching, and Universal-3.6 Pro Realtime streams 32 languages at $0.45 per hour with keyterm prompting.

Diarization is where AssemblyAI pulls ahead for multi-speaker products. It works on both pre-recorded and realtime models, and Speaker Identification can replace generic Speaker A and Speaker B labels with real names or roles, which saves you a post-processing step in call analytics or interview tools. On top of transcription, the Speech Understanding API and Guardrails add PII redaction, content moderation, and LLM-based extraction, and the LLM Gateway lets you run prompts over transcripts without wiring up another provider.

The free tier is generous for evaluation: up to 185 hours of pre-recorded or 333 hours of streaming audio, with no card. Streaming concurrency scales automatically with no extra fee. The trade-off is that the accurate models cover far fewer languages than Universal-2, so check your language list against the specific model before you commit.

Most developers know ElevenLabs for text-to-speech, but its Scribe speech-to-text models have become one of the best accuracy-per-dollar options in the category. Scribe v2 transcribes 90+ languages at $0.22 per hour of audio, with speaker diarization for up to 32 speakers, precise word-level timestamps, and entity detection across 65 entity types. Files can be up to 3 GB or 10 hours long, and multichannel mode handles up to 5 channels with a speaker ID per channel.

The feature that sets Scribe apart is audio-event tagging. Alongside words, the transcript marks non-speech sounds such as laughter and applause, which matters if you're building subtitle tools, podcast search, media archives, or training data pipelines where silence and reactions carry meaning. Keyterm prompting accepts up to 1,000 terms in batch for an extra $0.05 per hour, useful for brand names and technical vocabulary.

For live audio, Scribe v2 Realtime streams over WebSocket at about 150 ms latency in 90+ languages for $0.39 per hour, with word-level timestamps. The free and pay-as-you-go tier includes about 4.5 hours of Scribe v2, and the $6 Starter plan raises that to 27 hours. If your app also speaks back, keeping STT and TTS on one ElevenLabs account simplifies billing; see how it compares with another developer-focused voice platform in our Resemble AI vs ElevenLabs comparison.

Speechmatics has worked on speech recognition since 2006, and its 2026 line-up is aimed squarely at developers who handle messy, multilingual audio. Melia 1, launched in June 2026, transcribes 56 languages without being told which one to expect and switches languages mid-sentence, all for $0.12 per hour of pre-recorded audio. That's the lowest list price among the full-featured APIs in this guide.

What makes Speechmatics easy to budget is that diarization and word-level timings are included on every model rather than sold as add-ons. Enhanced, the highest-accuracy single-language model, costs $0.38 per hour batch and $0.40 per hour real-time and supports a custom dictionary. For voice agents, Linden 1 is a real-time-only model with built-in turn detection, listed at $0.15 per hour. There's also Oak 1 for clinical audio.

The other reason teams pick Speechmatics is deployment. The same models can run in Speechmatics' cloud, on-premises, or on-device, which matters for healthcare, government, and products that must work offline. New accounts get $100 in credits with no card. The main gap today is that Melia 1's code-switching isn't available in real time yet, so live multilingual transcription still means choosing a language per session on Enhanced or Standard.

Gladia is a Paris-based speech API that leans hard into multilingual audio. All of its 100+ languages support automatic language detection and mid-sentence code-switching, and the Solaria models are tuned for the kind of audio voice apps actually receive. Solaria-1 handles both real-time and async, while Solaria-3 is an async-only model built for messy call-center and sales recordings. Streaming returns partial transcripts in about 100 ms, and word-level timestamps are included on every plan.

Pricing is all-inclusive but starts higher than the competition: $0.61 per hour async and $0.75 per hour real-time on the pay-as-you-go Starter plan, with €50 in free credits. Commit upfront on Growth and rates drop to as low as $0.20 and $0.25 per hour, which puts Gladia in the same range as AssemblyAI and Deepgram for teams with predictable volume. Every feature and language is included in the price, so there are no surprise add-on lines.

Gladia is a natural fit for European teams: EU data residency, GDPR, HIPAA, SOC 2 Type II, and ISO 27001 come standard, and paid plans opt you out of model training. The main limitation for voice apps is diarization: it only works on pre-recorded audio, so live sessions need multichannel audio or an async re-run to label speakers.

The OpenAI API is the default for teams that already run their app on GPT models and want transcription on the same key. The recommended model, gpt-transcribe, costs about $0.0045 per minute (roughly $0.27 per hour) and accepts a free-form prompt, a keyword list, and expected languages to steer domain vocabulary. gpt-4o-mini-transcribe drops the price to about $0.18 per hour, and gpt-live-transcribe handles streaming audio over the Realtime API at about $0.017 per minute.

The catch for developers is fragmentation. OpenAI's own docs tell you to use a specialized model for each extra feature: gpt-4o-transcribe-diarize for speaker labels (with up to four known-speaker reference clips), and whisper-1 for word or segment timestamps or translation into English. Speaker labeling isn't supported in Realtime sessions, and file uploads are capped at 25 MB, so long recordings need compressing or chunking. If your app needs diarization, timestamps, and the best accuracy at once, you may end up calling two models on the same audio.

Where OpenAI shines is the pipeline after transcription. Passing a transcript to a GPT model for summaries, structured extraction, or tool calls happens in one account with one SDK, and the Realtime API and GPT-Live give you a route to full speech-to-speech agents later.

Groq isn't a speech company, but its LPU inference cloud hosts OpenAI's open Whisper models at prices nobody else matches. Whisper Large v3 Turbo costs $0.04 per hour of audio and Whisper Large v3 costs $0.111 per hour, through OpenAI-compatible transcription and translation endpoints. Groq lists real-time speed factors of 189x to 216x, so an hour-long file comes back in seconds rather than minutes.

For developers, that makes Groq a strong backend for batch jobs where you control the pipeline: transcribing user uploads, indexing video libraries, or generating training data. Because the endpoints mirror OpenAI's, switching from whisper-1 is usually a base-URL change. Word and segment timestamps are supported, and the same account gives you fast LLM inference for post-processing.

The limits are those of Whisper itself plus Groq's packaging. There's no speaker diarization and no streaming endpoint, so it can't power the live side of a voice agent. Uploads are capped at 25 MB on the free tier and 100 MB on the developer tier, every request is billed for at least 10 seconds, and only the first audio track of multi-track files is transcribed. Groq's own table lists Whisper Large v3 at 10.3% WER and Turbo at 12%, which is behind the commercial models above on hard audio.

Rev is best known for its consumer transcription service, but developers should look at Rev AI, its API platform, which runs on the same Reverb ASR models trained on Rev's large archive of human-verified transcripts. Reverb transcription costs $0.20 per hour for English and $0.30 per hour for 57 other languages, with async and streaming endpoints, speaker diarization, and add-ons for language identification, forced alignment ($0.003/min), translation, sentiment, summarization, and topic extraction.

The reason Rev earns a spot in a developer guide is the human fallback. The same API can route a file to human transcription for $1.99 per minute, which is useful for legal, medical, and media products where a small share of recordings must be near-perfect. You can run everything through Reverb by default and escalate only the files your users flag, without integrating a second vendor. Rev AI also offers hosted Whisper Large at $0.005 per minute.

The trade-offs are depth and pace. Rev AI's feature set and documentation are thinner than Deepgram's or AssemblyAI's, there's no turn detection for agents, and billing has a 15-second minimum per request. New accounts get free credits worth 5 hours of Reverb transcription.

Google Cloud Speech-to-Text makes sense mainly for teams already running on Google Cloud. The V2 API serves Chirp 3, Google's speech foundation model, across 100+ languages with automatic language detection, through streaming, synchronous (under one minute), and batch recognition. It plugs into IAM, Cloud Storage, US and EU regional endpoints, and your existing Google Cloud commit.

Pricing has two very different faces. Standard recognition costs $0.016 per minute (about $0.96 per hour) for the first 500,000 minutes a month, several times the specialist vendors, though it tiers down to $0.004 per minute above 2 million minutes. Dynamic batch, which processes audio at lower urgency, costs $0.003 per minute (about $0.18 per hour), making Google one of the cheapest options for clearing large archives.

For voice-app developers, the feature gaps matter more than price. Google's Chirp 3 docs state the model doesn't support word-level timestamps, diarization only works in batch and synchronous modes for 15 languages, and only about 20 languages are generally available while the rest are in preview. Setup also takes more work than a single API key: you'll create a project, recognizers, and service accounts before your first request.