What is Soniox Speech-to-Text?
Soniox Speech-to-Text is a speech recognition service that converts live or recorded audio into text through an API and a consumer app. It covers more than 60 languages with one model and handles speakers who switch language mid-sentence.
Developers send audio to two current models: stt-rt-v5 for streaming and stt-async-v5 for recorded files. Both return text with speaker labels and language identification, and both accept context such as names or product terms to guide recognition. The same account also offers text-to-speech with more than 200 voices and real-time translation across 3,600 language pairs.
Soniox is led by founder and CEO Klemen Simonic, with co-founder Ambroz Bizjak as chief architect, and its stated mission is helping the world understand every conversation. The homepage names Perplexity, Samsung, LG, Krisp and Fireflies.ai among its customers. Model releases continued through 2026, with the v5 generation replacing v4 at the end of June.
Key Features
Word-by-word streaming and automatic language switching carry Soniox, because together they let one model serve live multilingual calls without setup per language.
- Real-time streaming: The stt-rt-v5 model returns text word by word while someone is still speaking. Endpoint detection marks when a speaker has finished, which helps voice agents decide when to reply.
- Async file transcription: The stt-async-v5 model handles recorded files in bulk and can notify your system by webhook when a job finishes. Earlier release notes put the maximum audio length at five hours per file.
- Mixed-language speech: One model covers more than 60 languages and follows speakers who switch language mid-sentence. You do not need to pick a language first, which suits multilingual meetings and call centres.
- Speaker diarisation: Both v5 models separate speakers and label who said what in a conversation. The latest release notes describe this as rebuilt, with better handling of live, overlapping dialogue.
- Context and custom terms: You can pass names, product terms and domain vocabulary with each request to steer recognition. This matters for medical, legal and technical audio where generic models often misspell jargon.
- Built-in translation: Real-time translation runs inside the same API call as transcription, with 3,600 language pairs in one-way or two-way mode. The pricing page lists no extra charge for it.
Pros
Price is the headline advantage for Soniox: async transcription is listed at $0.10 an hour, with translation included at no extra charge.
- Low hourly cost: Async transcription is listed at $0.10 an hour and streaming at $0.12 an hour. That is a small fraction of many legacy cloud speech prices, so high-volume products can scale without large bills.
- Strong privacy defaults: Audio is not stored by default, and the docs state customer data never trains Soniox models. Stored async files are deleted after 30 days, which reduces risk for regulated teams.
- Compliance coverage: The company lists SOC 2 Type 2, ISO/IEC 27001:2022, HIPAA and GDPR. Compliance documents sit in the console, which shortens security reviews for enterprise buyers.
- One API for voice: Speech-to-text, text-to-speech and translation share one account and one pricing page. Teams building voice agents avoid stitching three vendors together and managing three sets of keys.
- App for non-developers: The Soniox App brings the same engine to iOS, Android, Mac, Windows and the web. Writers, students and field workers can dictate or record meetings without touching code.
Cons
Token-based billing is the main friction with Soniox, since budgets are easier to plan in minutes than in audio and text tokens.
- Token billing takes effort: Invoices are calculated in input and output tokens, not just minutes. You need the conversion guide, about 30,000 audio tokens per hour, to forecast costs accurately for long sessions.
- Benchmarks are self-published: The homepage calls Soniox the most accurate engine, but the evidence comes from its own comparison pages. Test your own accents, noise levels and jargon before trusting any vendor claim.
- Fast model turnover: The v3 models were retired in February 2026 and v4 in June 2026, with requests routed to newer versions. Teams with strict validation rules must retest output after each forced upgrade.
- Free app usage is vague: The free app plan runs on weekly credits, but the pricing page does not state how many minutes they cover. Heavy users will likely hit the cap and need the $19.99 Pro plan.
- No published free API tier: The API pricing page lists pay-as-you-go rates but no standing free allowance. Developers who want a long trial should check the console offer or contact sales first.
Who is Using Soniox Speech-to-Text?
Voice AI companies form the core of Soniox’s customer list, alongside device makers and meeting tools that need many languages from one engine.
- Voice agent builders: Companies such as Retell AI and Agora appear among customers. They use low-latency streaming so AI phone and video agents can listen and respond in near real time.
- Meeting and notes apps: Fireflies.ai, Krisp and Tana are named on the homepage. These products transcribe calls and notes in many languages and feed the text into summaries and search.
- Consumer device makers: Samsung, LG and Truecaller are listed as customers. Their use cases include voice features on phones, televisions and calling apps that must handle many languages.
- Individual professionals: Soniox markets its app to journalists, teachers, accountants and estate agents. They dictate drafts, record interviews and turn voice notes into searchable text on the move.
Pricing
Soniox bills API transcription by usage from $0.10 an hour, and its app runs from a free plan to $30 per user a month (official pricing pages, October 2026).
| Plan | Price | What you get |
|---|---|---|
| API async transcription | $0.10 per hour | File transcription; $1.50 per 1M input audio tokens, $3.50 per 1M input and output text tokens |
| API real-time transcription | $0.12 per hour | Streaming transcription; $2.00 per 1M input audio tokens, $4.00 per 1M input and output text tokens |
| API translation | No extra charge | Runs in the same call as transcription |
| API text-to-speech | About $0.70 per hour | $4.00 per 1M input text tokens and $21.50 per 1M output audio tokens |
| App Free | $0 | Smart Scribe, Translator and Voice Typing with weekly free credits |
| App Pro | $19.99 per month | Unlimited core features, priority processing and early access |
| App Business | $30 per user per month | Shared workspaces, seat management, data residency and compliance controls |
| Enterprise | Custom | Volume discounts and committed-use contracts |
The official guide equates one hour of audio to about 30,000 input audio tokens and about 15,000 output text tokens, which is how the hourly figures are derived. One directory lists the app from $20 a month, which matches the $19.99 Pro price once rounded. Annual app prices are said to be cheaper, but the figures were not shown.
Disclaimer: Soniox notes that enterprise and committed-use deals can differ from list rates, so confirm the figures in your console when you sign up.
What Makes Soniox Speech-to-Text Unique?
A single universal model is the core difference. Soniox does not ask you to choose a language before transcribing, and it keeps up when a speaker mixes two languages in one sentence. Many rival APIs still need a language code per request or separate models per region, which complicates global products.
The second difference is bundling. Translation runs inside the transcription call at no extra cost, and the same vendor supplies text-to-speech and a finished consumer app. Data residency is offered in the United States, the European Union, Japan and India, although the homepage describes the India region as limited early access.
Consent note: recording laws in many countries require every participant to agree before a call or meeting is recorded or transcribed, so tell people and get consent first. If you use the voice cloning feature, only clone a voice with the clear permission of its owner.
Compatibilities and Integrations
Voice agent frameworks are where Soniox plugs in most readily, with documented guides for LiveKit, Pipecat and Twilio alongside general AI developer kits.
- API and SDKs: Streaming and file transcription endpoints, with Python and Node.js examples in a public GitHub repository.
- Voice frameworks: Official guides cover LiveKit, Pipecat and Twilio for phone and real-time agent use.
- AI developer kits: Integrations are listed for the Vercel AI SDK, TanStack AI SDK and LangChain in Python and JavaScript.
- Automation: An n8n integration and async webhooks let you trigger workflows when transcripts are ready.
- App platforms: The Soniox App runs on iOS, Android, Mac, Windows and the web, with offline recording.
Soniox Speech-to-Text Tutorials
Your first streaming transcript needs an API key and one of the sample scripts from the official examples repository.
- Create an account in the Soniox Console and set up a project to generate an API key.
- Clone the soniox_examples repository and install the Python or Node.js dependencies.
- Run the async example on a short recorded file to check accuracy on your own audio.
- Switch to the real-time example and stream microphone audio to the stt-rt-v5 model.
- Add context terms such as names and product words, then turn on speaker labels or translation.
- Choose a data residency region and webhook settings before moving the project to production.
How We Rated It
Low hourly prices, strong privacy terms and true multilingual streaming put Soniox at 4.4 out of 5, with token billing and self-published benchmarks as the main caveats.
| Criterion | Score | Why |
|---|---|---|
| Accuracy and Reliability | 4.4/5 | Strong multilingual results reported |
| Ease of Use | 4.2/5 | Clear docs and SDK examples |
| Functionality and Features | 4.6/5 | STT, TTS and translation |
| Performance and Speed | 4.5/5 | Word-by-word streaming |
| Customization and Flexibility | 4.2/5 | Context terms and regions |
| Data Privacy and Security | 4.6/5 | No training, SOC 2, HIPAA |
| Support and Resources | 3.9/5 | Docs good, support unclear |
| Cost-Efficiency | 4.7/5 | From $0.10 per hour |
| Integration Capabilities | 4.4/5 | LiveKit, Pipecat, Twilio, n8n |
| Overall Score | 4.4/5 | Average of the nine criteria |
Summary
Soniox suits developers building voice agents, meeting tools or multilingual apps who want accurate streaming transcription at a low hourly rate. Its universal model, built-in translation and firm no-training policy make it a sensible default for global products. The main limitation is cost planning, because token billing and quick model retirements demand some engineering attention. Individuals can skip the API entirely and use the app on phone or desktop. Before committing, run your noisiest real recordings through both models and compare the results with your current provider.

Reviews
Write a Review