How to Transcribe Audio With ElevenLabs in 2026

To learn how to transcribe audio with ElevenLabs, upload an audio or video file to Speech to Text, select Scribe v2, review the speaker-labelled transcript, and export it in the format your workflow needs. The sharpest catch is not the upload process, which is straightforward, but the gap between a clean demo and a production transcript: the platform advertises very high accuracy, yet its own language table places Urdu and several other languages in a moderate word error rate band, while independent 2026 research shows that noise, accents, code-switching, and partial speech can produce severe and uneven degradation.

I approached this guide as an operational decision rather than a feature tour. A creator transcribing one interview has different needs from a newsroom processing confidential recordings, a developer streaming live calls, or a support team sending transcripts into an automated agent. The browser interface, Transcript Editor, batch REST endpoint, realtime WebSocket, pricing model, and privacy controls should therefore be assessed as separate parts of one speech pipeline, not as a single button labelled “transcribe”.

The sections below explain both the no-code and API routes, show where costs can expand, map export and integration options, and identify the failure modes that deserve a human check. All plan limits, file caps, model names, surcharges, language claims, and direct quotations were checked against current vendor documentation, primary announcements, or named 2026 research. Where a vendor benchmark does not reveal real-world performance, the limitation is stated rather than converted into a false universal accuracy score.

What ElevenLabs Actually Transcribes in 2026

ElevenLabs now treats speech to text as a first-class part of its audio platform. The current batch model is Scribe v2, designed for long-form transcription, subtitling, captioning, and archive processing. Scribe v2 Realtime is the streaming counterpart for meetings, agents, live captions, and interactive products. Scribe v1 was scheduled for removal on 9 July 2026, so current integrations should use scribe_v2 or scribe_v2_realtime rather than preserving an older model identifier.

The practical feature set is broader than plain text conversion. Scribe v2 can return word-level or character-level timing, predict language, label speakers, tag non-speech events such as laughter, remove filler words in no-verbatim mode, bias recognition towards supplied keyterms, detect or redact sensitive entities, identify known speakers from a workspace library, and separate multichannel recordings. Its documented model surface supports more than 90 languages and up to 32 predicted speakers. The API accepts audio and video uploads, hosted media through source_url, and source links from services including YouTube and TikTok, subject to accessibility and platform rules.

That positioning matters because ElevenLabs is not only a transcription vendor. Teams already using its text-to-speech, dubbing, voice isolation, Studio, or agents can keep more of the audio lifecycle inside one account. The wider ElevenLabs platform guide explains how those generation and editing tools fit together; transcription adds the inverse path, converting recorded speech back into searchable, editable data.

The trade-off is specialisation. ElevenLabs offers strong audio interoperability, but a meeting platform may provide richer calendar capture and action-item workflows, while a professional transcription service may offer contractual human accuracy. The right question is therefore not whether Scribe can produce text. It is whether its timing, speaker labels, privacy controls, exports, and pricing align with the downstream job.

“enabling businesses to build agents that can talk, type and take action”
Mati Staniszewski, Co-Founder and CEO of ElevenLabs, in a February 2026 company statement reported by Reuters

How to Transcribe Audio With ElevenLabs in the Browser

How to Transcribe Audio With ElevenLabs Step by Step

1. Open Speech to Text. Sign in to ElevenLabs, open the Speech to Text area, and choose Transcribe files. The web route is the fastest option for an individual file because it requires no API key or code.
2. Upload the source. Select an audio or video file. The documented browser cap is 3GB. Use a descriptive filename because it becomes the project label in the central transcript list.
3. Choose the language strategy. Leave language detection automatic for mixed or uncertain material, or specify the language when you know it and want to reduce ambiguity.
4. Configure transcript options. Enable speaker diarisation for interviews or panels, add subtitles when the output will be used for video, and provide important names or product terms when keyterm prompting is available in the selected surface.
5. Submit and review. After processing, open the file in the results pane. Clicking a word jumps playback to that point, which is useful for checking names, figures, and speaker changes.
6. Edit and export. Correct the transcript, rename speakers, adjust segments and timing, then export as plain text, JSON, HTML, SRT, or VTT.

A practical media workflow starts before the upload. Keep the original recording untouched, make a working copy, and record the date, consent status, language, and intended output. For podcasts, create a separate transcript master rather than editing the only subtitle file. The transcript supports search and repurposing, while subtitle cues have display constraints and should be treated as a separate deliverable. The AI podcast production guide shows why this distinction matters when an episode is later converted into show notes, clips, translated captions, and synthetic narration.

The browser route is best for occasional interviews, lectures, podcasts, internal videos, and research recordings where a person will inspect the result. It becomes less attractive when files arrive continuously, when jobs need metadata and audit IDs, or when the output must trigger another service automatically. Those cases justify the API route covered later.

Route	Best For	Core Strength	Primary Constraint
Browser Speech to Text	One-off files and editorial review	Fast upload, playback-linked editing, speaker relabelling	Manual steps and a 3GB upload cap
Batch API	Archives, apps, media pipelines	Webhooks, metadata, redaction, multichannel output	Engineering, authentication, retries, and usage billing
Realtime WebSocket	Live captions, meetings, agents	Partial and committed results at low latency	Audio streaming quality, session state, and concurrency
Human Productions	High-stakes final transcripts	Native-speaker review and optional verbatim treatment	Higher per-minute cost and turnaround

Preparing Audio Before Upload

Speech recognition quality begins with capture. A model cannot reliably reconstruct words that were never recorded clearly, and aggressive audio restoration can introduce artefacts that sound plausible to a listener but confuse an ASR system. Record the microphone close to the speaker, avoid clipping, reduce room echo, and isolate participants to separate channels where possible. For remote interviews, local tracks are usually safer than a single compressed call recording because each person retains a cleaner signal and fewer overlapping artefacts.

Use a lossless working file when available, but do not upsample a poor source merely to create a larger WAV. The API accepts major audio and video formats, and it provides a low-latency path for raw 16-bit PCM at 16kHz, mono, little-endian. Encoded MP3, M4A, AAC, FLAC, WAV, MP4, MOV, and similar mainstream formats are generally suitable, although exact codec compatibility should be checked during implementation. The API requires at least 100 milliseconds of audio and an uploaded file below 5GB; the browser documentation sets a lower 3GB cap.

For field reporting or interviews, the capture device remains part of the transcription system. The comparison in our voice recorder transcription guide is useful when choosing between a phone, dedicated recorder, meeting bot, and local device. The most important specification is not a marketing claim about “AI recording”. It is whether the device preserves a clean original, exposes standard file formats, and lets the user control cloud upload and deletion.

A Five-Minute Preflight Check

• Listen to the first minute and a busy middle section through headphones.
• Confirm that every expected speaker is audible and that the left and right channels are intentional.
• Write down uncommon names, acronyms, locations, medical terms, model numbers, and brand spellings for keyterm prompting.
• Decide whether filler words and false starts are evidence that must be preserved or clutter that can be removed.
• Confirm consent, retention, and access rules before sending personal or regulated recordings to a cloud service.

One original insight from production work is that transcription failures are often diagnosed too late. Teams usually inspect the text after processing, but the fastest quality gain comes from a sixty-second audio preflight before upload. That small step catches the wrong input channel, a muted participant, a corrupted export, or a music-heavy mix before credits are consumed and downstream automation begins.

Editing, Speaker Labels, and Export Formats

The Transcript Editor is more than a text box. It links words to playback, permits WYSIWYG corrections, allows exact segment timing changes, and supports splitting, merging, adding, deleting, and reassigning segments. Two adjacent segments can be merged only when they belong to the same speaker. Users can add or delete speakers, reorder tracks, change track colours, and recalculate word alignment after editing a segment. Deleting a speaker also deletes the speaker’s associated segments, so that action should be treated as destructive.

Speaker diarisation answers “who spoke when”, but the first pass usually returns generic labels rather than verified names. Rename speakers only after listening to multiple turns, particularly when voices are similar or participants interrupt one another. On an interview, verify the opening introduction, a mid-recording exchange, and the closing section. A label that is correct at the start can drift after overlapping speech or a long silence.

Exports should match the destination. Plain text is ideal for reading and search. JSON preserves structured timing and speaker data for software. HTML is convenient for web presentation. SRT and VTT are subtitle formats, but the subtitle editor maintains cues separately from the underlying transcript. Changing subtitle timing does not alter transcript segments, and editing transcript text does not automatically repair every subtitle cue. This separation prevents one deliverable from silently corrupting the other, but it also creates two review passes.

Creators who need transcript-based rough cuts may prefer a dedicated editor. Our Descript AI review covers the advantage of deleting spoken content by editing text, alongside the limits of its visual finishing tools. ElevenLabs is stronger when transcription sits beside voice generation, dubbing, subtitles, or an audio API; Descript is often stronger when the transcript is the primary editing interface for a podcast or talking-head video.

For publication, retain three versions: the untouched machine transcript, the corrected editorial transcript, and the final subtitle file. This audit trail makes it possible to trace an error, rerun a model, or prove what a human changed. It also prevents a common operational mistake in which a clean reading copy replaces the evidence-rich transcript that still contains timestamps, speaker labels, and audio-event markers.

Pricing, Credits, and the Real Cost of a Transcript

ElevenLabs presents two pricing surfaces that should not be conflated. ElevenCreative subscriptions provide pooled monthly credits shared across speech to text, text to speech, music, dubbing, voice changing, and other tools. The public Creative pricing page lists Free at $0 with 10,000 credits, Starter at $6 with 30,000, Creator at $22 with 121,000, Pro at $99 with 600,000, Scale at $299 with 1.8 million, and Business at $990 with 6 million. Speech to Text is estimated at 330 credits per minute, so every transcription reduces the balance available for other products.

ElevenAPI publishes direct dollar rates. At the time of verification, Scribe v2 batch was $0.22 per audio hour and Scribe v2 Realtime was $0.39 per audio hour. Keyterm prompting added $0.05 per hour, entity detection added $0.07 per hour, and API documentation described percentage surcharges of 20% and 30% respectively. Speaker-role detection adds 10%. Entity redaction requires entity detection and adds the same documented 30% surcharge. These representations can overlap in how the vendor displays costs, so procurement teams should validate the final invoice model inside their account before forecasting at scale.

Plan or Meter	Published Price	Included Allowance	Transcription Cost Risk
Creative Free	$0/month	10,000 pooled credits	No rollover; STT competes with every other product for credits
Creative Starter	$6/month	30,000 pooled credits	Commercial use included, but long recordings exhaust the pool quickly
Creative Creator	$22/month	121,000 pooled credits	Voice and transcription workloads share one balance
Creative Pro	$99/month	600,000 pooled credits	Higher capacity, still credit-based rather than unlimited
Creative Scale	$299/month	1.8M credits and 3 seats	Team usage can make ownership and chargeback unclear
Creative Business	$990/month	6M credits and 10 seats	Large shared pool, but not a substitute for usage governance
ElevenAPI Scribe v2	$0.22/audio hour	Usage-based	Optional keyterms, entities, redaction, roles, and channels increase cost
ElevenAPI Realtime	$0.39/audio hour	Usage-based	Live sessions add concurrency, network, and retry overhead

The hidden multiplier is multichannel billing. When use_multi_channel is enabled, ElevenLabs transcribes each channel independently and bills every channel for the full audio duration. A one-hour, four-channel recording therefore represents four billed audio hours before optional features. Another subtle rule applies when more than 100 keyterms are supplied: the API enforces a minimum billable duration of 20 seconds per request. Small clips can become disproportionately expensive if they are split too aggressively.

Meeting products create a different economic model because transcription, summaries, and integrations may be bundled by seat. The AI meeting notes guide maps that trade-off. ElevenLabs can be cheaper and more flexible as a raw speech engine, but a team must still build capture, calendars, permissions, summaries, action extraction, and delivery. The true comparison is total workflow cost, not cents per audio hour.

Building a Batch Transcription Workflow With the API

The batch endpoint is POST /v1/speech-to-text. A basic request sends a file and model_id=scribe_v2 as multipart form data. Authentication normally uses the xi-api-key header. Frontend clients can instead use a single-use token that expires after 15 minutes and can be consumed once, which avoids exposing a permanent API key in a browser. For remote media, source_url is the current field; cloud_storage_url is deprecated and should not be used in new builds.

curl -X POST “$ELEVENLABS_API_BASE/v1/speech-to-text” \
  -H “xi-api-key: $ELEVENLABS_API_KEY” \
  -F “file=@interview.wav” \
  -F “model_id=scribe_v2” \
  -F “diarize=true” \
  -F “timestamps_granularity=word” \
  -F ‘keyterms=[“Perplexity AI”,”Scribe v2″,”Karachi”]’

A production request should also carry an internal job identifier. When webhook=true, the endpoint returns early and delivers the result to configured speech-to-text webhooks. webhook_metadata can attach a small, two-level JSON object of up to 16KB, allowing the response to be correlated with a customer, media asset, consent record, or content-management job. Specify a webhook_id when only one destination should receive the result. Verify webhook signatures and make handlers idempotent so duplicate deliveries do not create duplicate articles or database records.

The response includes detected language, confidence, transcript text, and word objects with start, end, type, speaker_id, and sometimes channel_index. Non-speech content can appear as audio_event tokens. The API also supports additional export formats, transcript retrieval, and transcript deletion endpoints. Relevant integration surfaces include REST, webhooks, Python and JavaScript SDKs, the browser or Node client package, realtime WebSockets, source URLs, cloud storage through signed HTTPS links, speaker libraries, and downstream JSON processing.

Newsrooms and research teams should treat the transcript as evidence-bearing data. The AI tools for journalists guide explains why source preservation, timestamps, and auditability matter more than a polished summary. Store the original filename, content hash, model ID, request parameters, returned language probability, processing date, and reviewer identity. Without that metadata, a corrected transcript cannot be reproduced or defended.

A Safer Job State Model

• Received: The source exists, consent and access checks have passed, and no API request has started.
• Submitted: The API accepted the job and returned a request or transcription identifier.
• Processing: The webhook has not arrived; retries must not create a second billable job without a deliberate decision.
• Review Required: The transcript exists, but confidence, speaker labels, entities, or key terms need inspection.
• Approved: A human or defined quality gate has accepted the transcript for its intended use.
• Deleted: The application and vendor-side transcript have been removed according to retention policy.

Realtime Transcription With WebSockets

Scribe v2 Realtime is designed for audio that arrives continuously rather than as a finished file. The WebSocket accepts audio chunks and returns partial, committed, and timestamped transcription events. Applications can commit manually or rely on voice activity detection. ElevenLabs advertises latency around 150 milliseconds for the model, but end-to-end user experience also includes capture buffering, codec conversion, network travel, server scheduling, event handling, and rendering. A 150ms model claim is not the same as a 150ms caption appearing on a viewer’s screen.

Realtime keyterms support up to 50 entries of 20 characters each, compared with a much larger batch allowance. The April 2026 changelog also documented no-verbatim mode for realtime sessions, allowing filler words, false starts, and disfluencies to be removed. This is useful for readable live captions and agent input, but it can be inappropriate for legal records, qualitative research, or speech analysis where hesitation is meaningful.

Use raw 16kHz mono PCM when latency matters and the capture stack can provide it cleanly. For browser use, generate a single-use token rather than shipping an API key. Track session_started, partial transcript, committed transcript, errors, disconnects, and final closure. The client should be able to reconnect without repeating already committed text, and the server should preserve an ordered event log so later captions can be reconstructed.

“Voice is rapidly becoming the default interface between humans and technology, and enterprise deployments require a real-time platform”
Scott Stephenson, CEO and Co-Founder of Deepgram, in an IBM announcement, February 2026

Nick Holda, IBM Vice President of AI Technology Partnerships, described the same enterprise shift as “refining and modernizing their operations”. The point for developers is not that every application needs voice, but that a voice interface turns transcription errors into system inputs. If a transcript triggers a refund, medical instruction, account change, or database update, the acceptance threshold must be much stricter than it is for captions.

Mati Staniszewski offered a broader interface vision at Web Summit: “hopefully all our phones will go back in our pockets, and we can immerse ourselves in the real world around us, with voice as the mechanism that controls technology.” That future increases the value of low-latency recognition, but it also raises the cost of a confidently wrong transcript.

Accuracy, Language Coverage, and Benchmark Reality

ElevenLabs markets Scribe v2 as exceeding 98% accuracy and achieving a very low word error rate on industry benchmarks. Its documentation provides a more useful breakdown: English, French, German, Spanish, Japanese, Vietnamese, and many other languages are placed in an excellent band at 5% WER or below; Hindi and several other languages fall in a high-accuracy band above 5% and up to 10%; Arabic, Korean, Punjabi, Thai, and others sit in a good band above 10% and up to 20%; Urdu, Pashto, Somali, and several low-resource languages are listed in a moderate band above 25% and up to 50%.

That language table is more decision-relevant than one global accuracy number. A 2% error rate on clean English cannot be applied to a code-switched Urdu-English customer call recorded through a noisy mobile network. It also cannot predict whether the errors affect harmless function words or critical names, amounts, diagnoses, and product identifiers. Keyterm prompting helps with anticipated vocabulary, but it does not repair missing speech, crosstalk, or a damaged recording.

“models often hallucinate plausible but unspoken content under partial or degraded inputs”
Geeyang Tay and co-authors, WildASR research paper, March 2026

The WildASR study evaluated seven systems across environmental degradation, demographic shift, and linguistic diversity. Its clean, in-distribution reference averaged 5.7% error across models and languages, yet rankings and robustness changed sharply under real-world shifts. A separate 2026 Polish-language study found ElevenLabs Scribe performed best among the tested systems on its Polish general and medical datasets, which is encouraging but still domain-specific rather than proof of universal leadership.

Evidence Type	What It Tells You	What It Does Not Prove	Practical Use
Vendor accuracy claim	A headline view of model capability	Your accent, room, codec, vocabulary, or error severity	Use for initial screening only
Language WER band	Relative documented strength by language	Performance on your domain or code-switched audio	Prioritise pilot languages and review effort
Independent benchmark	Cross-model behaviour under a defined protocol	Every production environment	Identify failure modes and build a test set
Your labelled sample	Actual errors on your speakers and audio chain	Future drift after model or workflow changes	Set acceptance thresholds and regression tests

A strong evaluation uses at least thirty minutes of representative audio, manually transcribed by a competent reviewer. Score WER, but also count critical entity errors, speaker attribution errors, missed speech, invented speech, timestamp drift, and edit minutes per audio hour. For subtitles, test cue readability and synchronisation. For agents, test whether an error causes the wrong tool call. The most valuable metric may be cost per approved transcript, not raw WER.

Handling Multiple Speakers, Channels, and Sensitive Data

Diarisation and multichannel transcription solve related but different problems. Diarisation takes a mixed recording and predicts speaker turns. You can set num_speakers from 1 to 32 or let the model choose, and a threshold controls whether similar voices are merged or one voice is split into several labels. Multichannel mode assumes each channel contains one speaker and transcribes channels independently. It supports up to five channels and can return separate transcripts or a combined time-sorted result with channel_index values.

Use multichannel audio when the capture system genuinely isolates participants. Do not enable it merely because a stereo file exists. Many stereo recordings contain the same mixed programme on both channels; transcribing both would duplicate content and double the billed duration. Combined multichannel output also requires timestamps and does not support entity detection or redaction, so teams may need to choose between clean channel attribution and certain compliance features.

Sensitive-data tools deserve similarly precise interpretation. Scribe v2 can detect entity categories across personally identifiable information, protected health information, payment-card information, other entities, and offensive language. Redaction can replace content with a generic redacted token, an entity type, or an enumerated type. Detection and redaction are machine classifications, not legal guarantees. A missed account number remains exposed, while an over-broad redaction can remove evidence needed for a case or investigation.

ElevenLabs states that Scribe v2 supports SOC 2, ISO 27001, PCI DSS Level 1, GDPR, HIPAA pathways, data residency, and enterprise zero-retention controls. The API documentation says zero retention is available to enterprise customers when logging is disabled, and HIPAA use requires a Business Associate Agreement. Those terms must be confirmed in the organisation’s contract, workspace configuration, region, and data flow. A badge on a product page does not automatically make a particular implementation compliant.

For video teams, transcript privacy extends into captions, editing, and publishing. The AI video editor comparison shows how transcription now sits inside broader editing suites. Exported captions can leak names, private side conversations, or machine-invented phrases even when the original media remains restricted. Review the text as a publishable asset, not merely a technical derivative.

Common Failures and Performance Bottlenecks

Speaker Labels Drift or Multiply

Set an expected speaker count when it is known, then review interruptions and long silences. If one voice is split into several labels, raise the diarisation threshold cautiously. If two speakers are merged, lower it or use isolated channels. Renaming labels without checking the timeline can hide the error rather than fix it.

Names and Technical Terms Are Wrong

Supply a short, high-value keyterm list rather than dumping an entire glossary. Include spelling variants only when they are genuinely spoken. More than 100 batch keyterms introduce a minimum 20-second billable duration per request, and irrelevant terms can bias recognition towards words that were not said.

Long Jobs Feel Stalled

Scribe v2 internally chunks files longer than eight minutes and can process up to four segments concurrently. A fifteen-minute file uses two internal concurrent segments, while a two-hour file can use four. This improves throughput but does not eliminate queueing, upload time, webhook delays, or account concurrency limits. Display job states and preserve a retry decision rather than silently resubmitting.

The Transcript Is Polished but Unfaithful

No-verbatim mode removes filler words, false starts, and disfluencies. That can be ideal for readable captions and summaries, yet it can make a transcript unsuitable as a literal record. Decide the standard before processing and label cleaned transcripts clearly. Never present a readability-edited transcript as verbatim evidence.

Costs Exceed the Estimate

Check channel count, optional surcharges, repeated failed requests, and whether a Creative credit pool is also being spent on voice generation or dubbing. Separate testing, production, and human-review costs. A cheap API minute can still create an expensive approved transcript if editors must correct every speaker turn and proper noun.

Voice output and transcription are often confused because they share the ElevenLabs brand. The free AI voice generator comparison covers synthetic speech generation, which turns text into audio. Scribe performs the opposite operation. Credits, model choices, quality criteria, and legal risks differ, so procurement and analytics should not group them as one undifferentiated “voice AI” workload.

When ElevenLabs Is the Right Tool, and When It Is Not

ElevenLabs is a strong fit when transcription is part of a broader audio workflow. Media teams can move from recorded speech to transcripts, subtitles, dubbing, and synthetic voice in one ecosystem. Developers gain batch and realtime models, structured timestamps, diarisation, keyterms, entities, redaction, webhooks, and SDKs. Multilingual teams benefit from automatic language handling, although documented quality varies widely and must be tested by language.

It is less compelling when the real need is a complete meeting system with calendars, automatic attendance, action items, CRM synchronisation, and user-level governance. It may also be the wrong default for legal depositions, regulated medical records, or broadcast captions that require a contractual human accuracy process. ElevenLabs offers paid human Productions services from $2.00 per audio minute for transcript review, but availability is described as beta and limited to selected users, so it should not be assumed as an on-demand service for every account.

Choose a dedicated editing platform when transcript-based cutting is the central job. Choose a meeting assistant when capture and team memory are the central job. Choose an enterprise ASR specialist when self-hosting, custom acoustic models, detailed service-level terms, or deep contact-centre tuning dominate. Choose local Whisper or another open model when data must remain on controlled infrastructure and the team can accept engineering and evaluation work. ElevenLabs is most differentiated when high-quality speech recognition must connect naturally to other audio creation tools.

Use Case	ElevenLabs Fit	Why	Alternative to Consider
Podcast or interview transcript	Strong	Browser editing, speakers, timestamps, subtitles, audio ecosystem	Descript for transcript-led editing
Live captions or voice agent	Strong with testing	Realtime WebSocket, low latency, keyterms, no-verbatim mode	Deepgram or AssemblyAI for specialised voice stacks
Team meeting memory	Partial	Good raw transcription, but workflow layer must be built	Otter, Fireflies, Notta, or a suite-native assistant
Legal or medical final record	Conditional	Useful first draft and entity tools, not a substitute for required human assurance	Certified or contracted human transcription
Sensitive on-premises processing	Limited by deployment terms	Cloud-first public workflow; enterprise controls require contract review	Self-hosted or on-premises ASR
Multilingual media localisation	Strong	Transcription connects to subtitles, dubbing, and voice generation	Specialist localisation platform for complex review chains

The balanced recommendation is to pilot ElevenLabs on the exact recording path that will be used in production. Do not rank it first because its voice synthesis is well known, and do not reject it because a meeting competitor has more workflow features. Compare approved transcript cost, critical error rate, review time, privacy terms, integration effort, and downstream consequences. A transcription engine is successful only when the resulting text is reliable enough for its next action.

Our Content Testing Methodology

This guide used a documentation-led workflow because no authenticated ElevenLabs workspace or production API key was supplied. We reproduced the browser and API decision paths from the current Speech to Text product guide, Transcript Editor documentation, Scribe v2 capability pages, API request schema, realtime changelog, model deprecation notice, Creative pricing, and ElevenAPI pricing. We checked file caps, speaker and channel limits, keyterm constraints, timestamp options, export formats, surcharge rules, webhooks, retention controls, and model identifiers against those primary sources.

Accuracy claims were separated into three evidence classes. Vendor claims and language WER bands were reported as ElevenLabs documentation, not independent universal results. ElevenLabs’ published language WER table was used to explain the benchmark foundation. WildASR was used to show why clean benchmark performance may not transfer to degraded, accented, code-switched, or partial speech. We also reviewed a 2026 Polish comparison that reported strong Scribe performance on its tested data, while preserving the limitation that one language and domain cannot establish overall superiority.

Pricing was verified on 22 July 2026. The article distinguishes pooled ElevenCreative credits from dollar-denominated ElevenAPI rates and flags areas where vendor pages express the same option as both a direct amount and a percentage surcharge. Because account contracts, taxes, enterprise discounts, region settings, and billing interfaces can differ, readers should confirm their account’s live estimate before procurement. We did not claim hands-on accuracy numbers because an undocumented or unauthenticated test would not be reproducible.

This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.

Conclusion

ElevenLabs makes audio transcription easy to start and surprisingly deep to operationalise. A user can upload a file, let Scribe v2 detect language and speakers, correct the transcript against playback, and export text or subtitles without writing code. Developers can move further with webhooks, structured word timing, keyterms, entity controls, multichannel audio, realtime streaming, and the surrounding ElevenLabs audio stack.

The difficult decisions sit beyond the button. Browser and API limits differ. Creative credits and API rates describe different billing models. Every additional channel can multiply cost. A polished no-verbatim transcript may be unsuitable as evidence. Speaker labels can drift. Languages with excellent benchmark results can still fail in a noisy domain, while Urdu and several other languages carry materially weaker documented error bands. Independent research reinforces the need to test real recordings rather than inherit a vendor’s average.

The platform is therefore best treated as a capable transcription engine inside a controlled editorial or software process. Human review remains necessary where names, numbers, speaker identity, regulated information, or automated actions matter. Open questions for the rest of 2026 include how consistently Scribe v2 holds its accuracy across real-world accents and code-switching, how pricing is harmonised across Creative and API accounts, and whether enterprise retention and deployment options become more transparent to smaller teams.

Frequently Asked Questions

Does ElevenLabs Transcribe Audio?

Yes. ElevenLabs provides Speech to Text through Scribe v2 for uploaded audio or video and Scribe v2 Realtime for live audio. The browser interface supports editing, speaker labels, playback-linked words, and export. The API returns structured text, timestamps, speaker information, language data, and optional audio-event tags.

Is ElevenLabs Audio Transcription Free?

A Free ElevenCreative plan includes 10,000 pooled monthly credits, and Speech to Text consumes credits from that shared balance. The API has separate usage pricing. Free access is useful for testing short recordings, but it is not unlimited and does not include the same enterprise retention or contractual controls as a paid deployment.

How Accurate Is ElevenLabs Scribe v2?

ElevenLabs advertises over 98% accuracy and publishes language-specific WER bands. English and many European languages are listed at 5% WER or below, while Urdu and several low-resource languages are in a 25% to 50% band. Real accuracy depends on noise, accent, overlap, vocabulary, and recording quality.

Can ElevenLabs Transcribe Video Files?

Yes. The browser and API accept audio and video. The documented browser maximum is 3GB, while API file uploads must be below 5GB. The API also accepts source URLs for hosted media. Verify that you have permission to process and download any third-party content.

Can ElevenLabs Identify Multiple Speakers?

Yes. Speaker diarisation can label up to 32 predicted speakers. You can set an expected number or let the model estimate it. Multichannel mode supports up to five channels when each channel contains one speaker, but every channel is billed for the full recording duration.

Which Transcript Formats Can ElevenLabs Export?

The Transcript Editor exports plain text, JSON, HTML, SRT, and VTT. SRT and VTT are designed for captions and subtitles. Transcript segments and subtitle cues are maintained separately, so editing one does not automatically update the other.

Does ElevenLabs Offer a Speech-to-Text API?

Yes. The batch REST endpoint accepts multipart uploads or source URLs, while the realtime service uses WebSockets. Integrations can use API keys, single-use frontend tokens, webhooks, Python and JavaScript SDKs, structured JSON responses, transcript retrieval, and deletion endpoints.

Is ElevenLabs Suitable for Confidential Recordings?

It can be, but suitability depends on the contract and configuration. ElevenLabs documents enterprise zero retention, data residency, security certifications, and a HIPAA pathway that requires a Business Associate Agreement. Teams should verify retention, access, region, deletion, and incident obligations before processing sensitive audio.

References

ElevenLabs. (2026). Transcription: Speech to Text documentation.

ElevenLabs. (2026). Create transcript API reference.

ElevenLabs. (2026). ElevenAPI pricing.

ElevenLabs. (2026, January 9). Introducing Scribe v2.

Reuters. (2026, February 4). ElevenLabs secures $11 billion valuation in latest funding round.

IBM. (2026, February 24). Deepgram and IBM introduce advanced voice capabilities for enterprise AI.

Tay, G., Ma, W., Lee, J., et al. (2026). Back to basics: Revisiting ASR in the age of voice agents.

Pietroń, M., Piórkowski, S., Faber, K., et al. (2026). Quality of automatic speech recognition: Polish language case study from Wav2Vec to Scribe ElevenLabs.

Google Search Central. (2026). Spam policies for Google web search.