ElevenLabs Speech to Text: Models, Benchmarks, Pricing and When to Use It

Multilingual and Indic voice agents — the only major provider combining sub-150ms streaming, 11 Indian languages, and India data residency. Models: Scribe v2 (batch), Scribe v2 Realtime (streaming), Scribe v1 (legacy). Latency: 3.6% WER at 0.14s after end of speech — the lowest error rate of any streaming model measured (Artificial Analysis, July 2026 — independent). Price: Scribe $0.22/hr batch, Scribe Realtime $0.39/hr, cut roughly 45% in May 2026 (ElevenLabs API pricing page, verified July 2026 — check the vendor's pricing page before relying on this). Honest limitation: No self-hosting, no on-premises, no air-gapped option. ElevenLabs is cloud-only. If your compliance posture requires the model to run inside your perimeter, this is a hard stop regardless of how good the accuracy is.

What is ElevenLabs speech to text?

ElevenLabs speech to text is the transcription half of a company better known for voice synthesis. Scribe v1 launched in February 2025; Scribe v2 Realtime and Scribe v2 batch both shipped across late 2025 and January 2026, splitting the product into a latency-optimised streaming model and an accuracy-optimised long-form model.

The architectural claim behind the streaming model is negative latency prediction — anticipating upcoming words and punctuation from context rather than waiting for the audio that confirms them. Combined with text conditioning for long segments and manual commit controls, this is how a model returns transcripts in roughly 140ms without the usual accuracy penalty.

What makes it interesting for Indian teams specifically is a detail buried in the release notes. Scribe v2 transcribes English words inside Indic-language audio in Latin script, rather than transliterating them into Devanagari or Tamil script. Most systems get this wrong, which produces technically-correct-but-unusable transcripts for bilingual professional speech. It works with no language configuration.

Which ElevenLabs STT models are available?

Two current models, one legacy.

Scribe v2 Realtime

The streaming model, delivering live transcription in under 150 milliseconds across 90+ languages. ElevenLabs reports 93.5% accuracy on the FLEURS benchmark across 30 European and Asian languages.

Features include voice activity detection, manual commit control over when transcripts finalise, and negative latency prediction. It integrates directly with ElevenLabs Agents.

Use it for conversational agents, live captioning, and real-time compliance monitoring.

Scribe v2 (batch)

Released 9 January 2026 and optimised for long and complex recordings rather than latency. It carries the heavier feature set: speaker diarization for up to 48 distinct speakers, entity detection across 56 PII categories, dynamic audio tagging for non-speech events, and keyterm prompting expanded from 100 to 1,000 terms in March 2026.

A March 2026 update added PII auto-redaction during transcription — before storage, not after — and a No Verbatim mode that strips filler words automatically.

Use it for podcasts, legal transcripts, medical dictation, and subtitle pipelines.

Scribe v1

The original February 2025 model. Superseded, but still referenced in older integrations.

A note on documentation

ElevenLabs publishes a per-language WER breakdown in its transcription docs, grouped into accuracy tiers. That is unusually transparent — most vendors publish one averaged number and leave you to discover which languages drag it down. Check the tier your language sits in before committing.

What features does ElevenLabs STT actually give you?

Accuracy and formatting. Keyterm prompting up to 1,000 terms on batch. Entity detection across 56 PII categories. Dynamic audio tagging for laughter, music and other non-speech events. No Verbatim mode for automatic disfluency removal.

Languages and code-switching. 90+ languages, including 11 Indian ones: Hindi, Tamil, Malayalam, Telugu, Gujarati, Kannada, Odia, Bengali, Marathi, Punjabi and Sindhi. Automatic language detection with multi-language files handled in a single request — no manual segmentation. Mixed-script handling keeps English in Latin script inside Indic audio.

Real-time behaviour. Sub-150ms transcription with predictive output, built-in VAD, and developer control over transcript finalisation. Speaker diarization is batch-only — you cannot diarize a live stream, which rules Scribe out for live multi-speaker meeting attribution.

Compliance and data residency. SOC 2, ISO 27001, PCI DSS L1, HIPAA (BAA via sales), and GDPR. EU and India data residency, plus zero retention mode. ElevenLabs also maintains an India-specific offering integrating with Ozonetel, Exotel and Plivo for inbound and outbound calling.

Developer experience. REST and WebSocket. Multichannel transcription across up to 5 channels, each assigned its own speaker ID, with combined duration under 10 hours. Async results can post to webhooks. API usage bills in US dollars; the web UI bills in credits at roughly 330 credits per minute — two different systems, and mixing them up is the usual source of confusion.

Benchmarks — how accurate is ElevenLabs really?

On the one genuinely independent streaming comparison available, ElevenLabs currently posts the lowest word error rate of any model measured.

BenchmarkMetricScoreSourceVendor or independentAA-WER StreamingWER / time after end of speech3.6% @ 0.14sArtificial Analysis, July 2026IndependentFLEURS, 30 European and Asian languagesAccuracy93.5%ElevenLabs, 2025VendorCode-switching, 18 languagesNormalised WER8.77%AssemblyAI, 2026CompetitorPer-language WER breakdownWER by language tierPublished in docsElevenLabs, 2026Vendor

The full independent streaming table puts that number in context:

ModelWERTime after end of speechElevenLabs Scribe v2 Realtime3.6%0.14sCartesia Ink-23.7%0.09sAssemblyAI Universal-3.5 Pro Realtime4.1%0.44sDeepgram Nova-3 Realtime6.6%0.07sDeepgram Flux7.4%0.02s

Source: Artificial Analysis, July 2026 — independent.

ElevenLabs wins on accuracy and sits second on latency. Only Cartesia is meaningfully faster at comparable accuracy. Deepgram Flux is roughly 7× faster to first final but at more than double the error rate — a genuine engineering trade-off, not a quality gap.

Note the third row of the benchmark table above. The 8.77% code-switching figure comes from AssemblyAI's published comparison, where they report beating Scribe v2 at 7.69%. A competitor's benchmark of your model is not independent evidence, in either direction. We include it because you will encounter it, and because two vendors claiming code-switching leadership against each other tells you the honest answer is "test both."

What these numbers do not tell you. FLEURS is read speech recorded in clean conditions. Telephony audio is 8kHz narrowband with compression artefacts and packet loss. A model at 3.6% WER on AA-WER's dataset can behave very differently on a GSM call from a Bengaluru street.

That gap widens for Indic audio specifically. There is no independent Indic ASR leaderboard — none of FLEURS, AA-WER or the Hugging Face Open ASR Leaderboard covers Indian languages meaningfully. So every Indic accuracy claim from every vendor, ElevenLabs included, is currently unfalsifiable from the outside.

WER on your own call recordings is the only number that decides anything.

How much does ElevenLabs cost?

Rates from the ElevenLabs API pricing page, verified July 2026 — check the vendor's pricing page before relying on this.

ProductPriceBilling unitScribe (batch)$0.22/hrPer hour of audioScribe Realtime$0.39/hrPer hour of audio

Prices dropped sharply in May 2026. ElevenLabs cut speech-to-text pricing by up to 45%, taking the batch rate from $0.40/hr to $0.22/hr, and introduced pay-as-you-go across the API. If you costed ElevenLabs before May 2026 and rejected it on price, re-run the numbers.

The headline rate is not the bill. Add-ons price separately:

Add-onPriceEffect on baseEntity detection (API only)$0.07/hr+32% on batchKeyterm prompting$0.05/hr+23% on batch

Enable both on batch and your effective rate is roughly $0.34/hr — about 55% above the headline. That is still competitive, but it is not $0.22.

Two pricing inconsistencies to be aware of. ElevenLabs' own speech-to-text marketing page still quotes "starting from $0.40 per hour," which the API pricing page contradicts at $0.22. Separately, at least one detailed third-party guide quotes Scribe v2 Realtime at $0.28/hr against the vendor page's $0.39/hr. Take the API pricing page as authoritative and confirm your rate in the console before modelling volume.

The UI and API also bill differently — credits versus dollars — so a cost estimate built from web-app usage will not match your API invoice.

What is ElevenLabs best used for?

Multilingual customer support for an Indian D2C brand — one model covering Hindi, Tamil, Telugu and Kannada with automatic detection, so you are not routing calls to per-language models. Loan-servicing calls for an Indian NBFC under DPDP obligations — India data residency plus zero retention mode, with Ozonetel or Exotel already integrated for telephony. Live compliance monitoring for a brokerage — sub-150ms streaming with entity detection flagging PII as it is spoken. Podcast and long-form media captioning for a publisher — Scribe v2 batch with 48-speaker diarization and audio tagging for non-speech events. Clinical documentation for a hospital group — HIPAA BAA and PII auto-redaction before storage, though on cloud infrastructure only. Subtitle pipelines for an OTT platform — 90+ languages with multi-language files handled in one request, no manual segmentation.

When should you use ElevenLabs?

Your agent operates in Indian languages beyond Hindi and you need it in real time. Eleven Indic languages with sub-150ms streaming is a combination nobody else currently offers. Accuracy is the binding constraint and you can absorb ~140ms. On the only independent streaming benchmark available, nothing is more accurate. You need India or EU data residency without building it yourself. You are already running ElevenLabs TTS and want transcription under the same contract, key and compliance posture.

When should you not use ElevenLabs?

You need on-premises, air-gapped or self-hosted deployment. ElevenLabs is cloud-only, full stop. Use Speechmatics, which offers on-prem, on-device and air-gapped deployment across 56+ languages, or Mistral Voxtral Realtime under Apache 2.0 if you want to run the weights yourself. Turn-taking latency is your bottleneck. 0.14s is good, but Deepgram Flux posts 0.02s with model-integrated end-of-turn detection. If your agent feels sluggish and you can tolerate higher WER, that is the swap. You need live speaker diarization. Scribe's diarization is batch-only. Use AssemblyAI, whose Universal-3.5 Pro Realtime ships self-correcting speaker labels on streaming audio at $0.12/hr. You are transcribing large archives and cost dominates. $0.22/hr is fair but not cheapest. Use Groq at $0.04/hr for whisper-large-v3-turbo, roughly 5× cheaper for offline work.

How does ElevenLabs compare to the alternatives?

ElevenLabsDeepgramAssemblyAISarvamFlagship streaming modelScribe v2 RealtimeFlux EnglishUniversal-3.5 Pro RealtimeSaaras v3Streaming price$0.39/hr$0.39/hr (promo)$0.45/hr (or $0.15/hr English)₹30/hr ($0.36/hr)Independent WER / latency3.6% @ 0.14s7.4% @ 0.02s4.1% @ 0.44sNot independently benchmarkedStreaming languages90+45+1822 Indian + EnglishIndian languages11Hindi + 5 monolingualHindi only22India data residencyYesNoNoIndia-basedLive diarizationNo — batch onlyYes, $0.12/hrYes, $0.12/hrYesSelf-hostNoYesYesNo

WER and latency from Artificial Analysis, July 2026 (independent). Prices verified July 2026 from vendor pages.

Worth stating plainly: for Indic real-time voice agents, ElevenLabs and Sarvam are the two serious options, and they win on different axes. Sarvam covers 22 Indian languages and was trained on Indian speech from the ground up. ElevenLabs covers 11 with better independently-verified accuracy on non-Indic audio and a broader global footprint. If your traffic is majority Indic, test Sarvam first.

How do you use ElevenLabs on Telenow?

ElevenLabs is live on Telenow as a selectable STT component. In the agent builder, open the Speech-to-Text step and select ElevenLabs from the ElevenLabs STT integration page. Choose Scribe v2 Realtime for live agents; Scribe v2 batch is the right choice for post-call transcription pipelines. Enable entity detection only where you need PII flagging, since it adds $0.07/hr.

If you are already running ElevenLabs for TTS, note that Telenow still itemises the two separately. STT and TTS appear as distinct cost lines even under one vendor, so you can see exactly what transcription costs you and swap either half independently.

Your cost line breaks out properly. A one-hour Tamil support call shows ElevenLabs STT at $0.39/hr, your chosen LLM at its own rate, TTS separately, and Indian carrier minutes separately. No blended per-minute figure hiding a markup. Compare against the full component pricing breakdown.

Swapping later does not mean rebuilding. If an audit requires on-premises processing, move the STT component to Speechmatics. If your Indic traffic grows past the 11 supported languages, move to Sarvam. Prompts, tools, call logic and carrier configuration stay exactly as they are. Browse the full STT provider comparison or the voice AI platform overview.

Build on Scribe v2 Realtime and measure WER on your own Indic call recordings. Given no independent Indic benchmark exists, that is not just the best test — it is the only one. What is the difference between Scribe v2 and Scribe v2 Realtime? Scribe v2 Realtime is optimised for live streaming at sub-150ms latency for voice agents and captioning. Scribe v2 batch is optimised for long-form recordings, adding 48-speaker diarization, entity detection across 56 PII categories, audio tagging and keyterm prompting up to 1,000 terms.

3. JSON-LD

json[ { "@context": "https://schema.org", "@type": "Article", "headline": "ElevenLabs Speech to Text: Models, Benchmarks, Pricing and When to Use It", "description": "ElevenLabs speech to text explained: Scribe v2 and v2 Realtime, verified pricing, independent WER benchmarks, 11 Indian languages and honest limitations.", "url": "https://telenow.ai/blog/elevenlabs-speech-to-text", "datePublished": "2026-07-18", "dateModified": "2026-07-18", "author": { "@type": "Organization", "name": "Telenow", "url": "https://telenow.ai" }, "publisher": { "@type": "Organization", "name": "Telenow", "url": "https://telenow.ai", "logo": { "@type": "ImageObject", "url": "https://telenow.ai/logo.png" } }, "mainEntityOfPage": { "@type": "WebPage", "@id": "https://telenow.ai/blog/elevenlabs-speech-to-text" }, "about": [ { "@type": "SoftwareApplication", "name": "ElevenLabs Scribe" }, { "@type": "Thing", "name": "Speech recognition" } ] }, { "@context": "https://schema.org", "@type": "SoftwareApplication", "name": "ElevenLabs Scribe Speech-to-Text API", "applicationCategory": "DeveloperApplication", "applicationSubCategory": "Speech Recognition API", "operatingSystem": "Cloud", "url": "https://elevenlabs.io/speech-to-text", "softwareVersion": "Scribe v2, Scribe v2 Realtime", "featureList": [ "Sub-150ms real-time streaming transcription", "90+ languages including 11 Indian languages", "Negative latency prediction", "Speaker diarization up to 48 speakers (batch)", "Entity detection across 56 PII categories", "PII auto-redaction before storage", "Keyterm prompting up to 1,000 terms", "EU and India data residency with zero retention mode" ], "offers": [ { "@type": "Offer", "price": "0.22", "priceCurrency": "USD", "description": "Scribe batch transcription, per hour of audio, verified July 2026", "url": "https://elevenlabs.io/pricing/api" }, { "@type": "Offer", "price": "0.39", "priceCurrency": "USD", "description": "Scribe Realtime streaming transcription, per hour of audio, verified July 2026", "url": "https://elevenlabs.io/pricing/api" } ], "provider": { "@type": "Organization", "name": "ElevenLabs, Inc." } }, { "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "How much does ElevenLabs speech to text cost per hour?", "acceptedAnswer": { "@type": "Answer", "text": "Scribe batch transcription costs $0.22 per hour of audio and Scribe Realtime costs $0.39 per hour, verified July 2026. Entity detection adds $0.07 per hour and keyterm prompting adds $0.05 per hour. ElevenLabs reduced speech-to-text pricing by up to 45% in May 2026." } }, { "@type": "Question", "name": "Which Indian languages does ElevenLabs Scribe support?", "acceptedAnswer": { "@type": "Answer", "text": "Scribe v2 Realtime supports 11 Indian languages: Hindi, Tamil, Malayalam, Telugu, Gujarati, Kannada, Odia, Bengali, Marathi, Punjabi and Sindhi. It also transcribes English words inside Indic audio in Latin script rather than transliterating them, which matters for bilingual professional speech." } }, { "@type": "Question", "name": "Is ElevenLabs Scribe more accurate than Deepgram?", "acceptedAnswer": { "@type": "Answer", "text": "Yes, on the available independent data. Artificial Analysis measured Scribe v2 Realtime at 3.6% WER in July 2026 against Deepgram Flux at 7.4% and Nova-3 Realtime at 6.6%. Deepgram is substantially faster, at 0.02s versus 0.14s after end of speech." } }, { "@type": "Question", "name": "Can ElevenLabs speech to text be self-hosted?", "acceptedAnswer": { "@type": "Answer", "text": "No. ElevenLabs is cloud-only, with no on-premises, air-gapped or self-hosted deployment option. It offers EU and India data residency plus zero retention mode as alternatives, but the model always runs on ElevenLabs infrastructure rather than inside your perimeter." } }, { "@type": "Question", "name": "Does ElevenLabs Scribe support speaker diarization?", "acceptedAnswer": { "@type": "Answer", "text": "Yes, but only for batch transcription, where it identifies up to 48 distinct speakers with timestamps. Real-time speaker diarization is not available on Scribe v2 Realtime, so live multi-speaker attribution requires a different provider." } }, { "@type": "Question", "name": "What is the latency of ElevenLabs Scribe v2 Realtime?", "acceptedAnswer": { "@type": "Answer", "text": "ElevenLabs states under 150 milliseconds. Artificial Analysis independently measured 0.14 seconds after end of speech in July 2026. Third-party testing reports p95 around 250ms and p99 around 400ms, so 150ms is a median rather than a worst case." } }, { "@type": "Question", "name": "Does ElevenLabs offer India data residency?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. ElevenLabs provides India data residency alongside EU residency and zero retention mode, and holds SOC 2, ISO 27001, PCI DSS L1, HIPAA and GDPR compliance. It also integrates with Indian telephony providers including Ozonetel, Exotel and Plivo." } }, { "@type": "Question", "name": "What is the difference between Scribe v2 and Scribe v2 Realtime?", "acceptedAnswer": { "@type": "Answer", "text": "Scribe v2 Realtime is optimised for live streaming at sub-150ms latency for voice agents and captioning. Scribe v2 batch is optimised for long-form recordings, adding 48-speaker diarization, entity detection across 56 PII categories, audio tagging and keyterm prompting up to 1,000 terms." } } ] } ]

4. Internal links used

Anchor textDestinationElevenLabs STT integration page/voice-ai/stt/elevenlabsfull STT provider comparison/voice-ai/sttvoice AI platform overview/voice-aifull component pricing breakdown/pricingSpeechmatics (in "when not to use")/blog/speechmatics-speech-to-textMistral (in "when not to use")/blog/mistral-voxtral-speech-to-textDeepgram (in "when not to use")/blog/deepgram-speech-to-textAssemblyAI (in "when not to use")/blog/assemblyai-speech-to-textGroq (in "when not to use")/blog/groq-whisper-speech-to-textSarvam (comparison section)/blog/sarvam-speech-to-text

External links: elevenlabs.io/pricing/api · elevenlabs.io/docs/overview/capabilities/speech-to-text · elevenlabs.io/blog/introducing-scribe-v2 · elevenlabs.io/blog/weve-lowered-api-agents-pricing-and-introduced-pay-as-you-go · elevenlabs.io/india · artificialanalysis.ai/speech-to-text

5. Image concepts

Accuracy-versus-latency scatter plot. The five Artificial Analysis July 2026 points — ElevenLabs, Cartesia, AssemblyAI, Deepgram Nova-3 RT, Deepgram Flux — WER on Y, time-after-end-of-speech on X. ElevenLabs sits lowest on the accuracy axis, second-left on latency. Alt: "ElevenLabs speech to text accuracy versus latency compared with Deepgram, AssemblyAI and Cartesia on Artificial Analysis benchmarks" Mixed-script Indic transcript comparison. Two panels of the same Hinglish utterance: a typical system transliterating "dashboard" into Devanagari, versus Scribe v2 keeping it in Latin script. Alt: "ElevenLabs speech to text keeping English words in Latin script within Hindi audio instead of transliterating them" Indian-language coverage map. India shaded by the 11 supported languages, with a data-residency badge and the Ozonetel, Exotel and Plivo telephony integrations noted alongside. Alt: "ElevenLabs speech to text coverage of 11 Indian languages with India data residency for voice agents"