> ## Documentation Index
> Fetch the complete documentation index at: https://api-docs.ollang.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Speech-to-Text Providers
> Compare the transcription engines available in Ollang Workflows — accuracy, language coverage, code-switching behavior, and which content type each one suits.
Speech-to-text runs in the `transcribe` step. It produces the timecoded, speaker-attributed transcript that Closed Captions ship directly, that Subtitle Orders translate, and that AI Dubbing turns into a dubbing script.
Transcription is the step where errors do the most damage, because everything downstream inherits them. A mistranscribed proper noun becomes a mistranslated proper noun and then a mispronounced one.
## At a glance
| Provider | Provider tag | Languages | Best for | Status |
| ----------------------------------------------- | ------------------------------ | ------------ | --------------------------------------------------------------- | ------------------------ |
| [Agentic STT](#agentic-stt) | `ollangAITranscribe` | — | Complex multi-speaker media, noisy audio, mixed languages | **Default** |
| [Agentic STT-1.2](#agentic-stt-12) | `ollang-stt-1.2-exp` | — | Same workloads, next-generation agent orchestration | Experimental |
| [Visual STT](#visual-stt) | `agentic-visual-transcription` | — | Silent video, on-screen text, kinetic typography | Experimental |
| [AssemblyAI](#assemblyai) | `assemblyai` | 99 | General-purpose transcription with speech intelligence | Available |
| [ElevenLabs Scribe v2](#elevenlabs-scribe-v2) | `elevenlabs-scribe-v2` | 90+ | Under-resourced languages, 32-speaker diarization, audio events | Available |
| [Speechmatics](#speechmatics) | `speechmatics` | 70+ | Accented and dialectal speech, automatic language switching | Available |
| [Speechmatics with LLM](#speechmatics-with-llm) | `speechmatics-llm` | 70+ | Same, plus custom guidelines and instructions | Available |
| [AWS Transcribe](#aws-transcribe) | `aws` | 100+ locales | AWS-aligned enterprise pipelines, nine English locales | Available |
| [Whisper](#whisper) | `whisper` | 90+ | — | Not currently selectable |
The Ollang agentic providers orchestrate several underlying models rather than exposing a fixed language list, so no single coverage figure applies to them. The counts shown are the vendor-published figures for the single-model providers. AWS is listed in locales rather than languages — nine of its entries are English variants and three are Spanish.
AssemblyAI and ElevenLabs Scribe v2 are the single-model workhorses of this catalog — the natural first stop for a large backlog of reasonable-quality audio. Reserve Agentic STT for content where the audio is genuinely difficult, since that is where a multi-agent pass earns its extra processing.
***
## Agentic STT
Ollang's own multi-agent transcription workflow, and the default for every transcription step.
Rather than sending audio to a single ASR model, Agentic STT runs a collaborative agent pass that reconciles competing hypotheses, resolves speaker attribution across the whole timeline, and uses surrounding context — including non-speech audio cues — to decide between plausible readings.
**Best for** — video and audio with multiple language streams, several speakers, background noise, or diverse accents; anything where a single-pass ASR model produces a transcript that reads plausibly but is wrong.
**Strengths**
* Superior handling of noisy audio, overlapping speech, and accent diversity compared with single-model ASR.
* Contextual interpretation of audio cues, which matters for localization workflows where the transcript becomes a dubbing script.
* Consistent speaker attribution across long-form content.
**Trade-offs**
* A multi-agent pass does more work than a single model call, which shows up in turnaround on large batches. On clean, single-speaker studio audio it buys very little over the single-model providers.
**In Ollang** — `ollangAITranscribe` · Closed Captions, Subtitles, AI Dubbing · Default.
***
## Agentic STT-1.2
The next iteration of the agentic transcription workflow, available for evaluation alongside the current default.
It targets the same content profile as Agentic STT — multi-stream, noisy, accent-diverse media — with revised agent orchestration.
**Best for** — teams that want to benchmark the next generation against their existing Agentic STT baseline before it becomes the default.
Experimental. Behavior may change between releases. Run it in a benchmarking Folder alongside Agentic STT rather than switching a production Global Workflow to it.
**In Ollang** — `ollang-stt-1.2-exp` · Closed Captions, Subtitles, AI Dubbing · Experimental.
***
## Visual STT
A different kind of transcription: instead of listening, it reads. Visual STT extracts on-screen text from video frames and turns it into editable, timecoded text content.
**Best for** — silent video, social-first content where the message lives in burned-in captions, product demos with UI text, title cards, lower thirds, and kinetic typography.
**Strengths**
* Recovers content that audio-only transcription cannot see at all.
* Produces editable text, so on-screen copy can be localized rather than left in the source language.
**Trade-offs**
* It transcribes what is *displayed*, not what is *said*. For content that has both meaningful speech and meaningful on-screen text, you will want a separate audio transcription pass as well.
* Preview quality. Review output before it reaches a deliverable.
**In Ollang** — `agentic-visual-transcription` · Closed Captions, Subtitles, AI Dubbing · Experimental.
***
## AssemblyAI
A general-purpose speech AI platform, and the broadest language coverage of any transcription provider in the catalog.
AssemblyAI's model line splits along a breadth-versus-accuracy axis:
| Model | Languages | Notable |
| --------------------- | ----------------------------------- | --------------------------------------------------------------------------- |
| **Universal-2** | 99 | Broad coverage for general transcription |
| **Universal-3.5 Pro** | 21 (18 base plus regional variants) | Highest accuracy tier, with native code-switching including English–Spanish |
**Best for** — batch transcription of podcasts, meetings, interviews, and broadcast content, where audio quality is reasonable and language breadth matters.
**Strengths**
* **99 languages** on Universal-2 — the highest published language count in this catalog.
* Native code-switching on the Pro tier, for speakers who move between languages mid-sentence.
* Custom vocabulary support, which measurably improves handling of product names, people names, and domain jargon.
* Speaker diarization and speech intelligence — sentiment, topic detection, PII redaction — produced alongside the transcript rather than in a separate pass.
**Trade-offs**
* The highest-accuracy tier covers 21 languages, not the full 99. Confirm which tier serves your language before assuming top-tier accuracy.
* Accuracy drops on heavily accented speech and messy conversational audio, as it does for every single-pass model. That is where the agentic providers earn their premium.
**In Ollang** — `assemblyai` · Closed Captions, Subtitles, AI Dubbing.
***
## ElevenLabs Scribe v2
ElevenLabs' transcription model, and the most transparent provider in the catalog about where it is and is not accurate.
Scribe v2 transcribes **90+ languages**, and ElevenLabs publishes a **per-language accuracy tier** for every one of them — from *Excellent* (5% word error rate or below) through *Moderate* (25–50% WER). That published breakdown is the most useful thing about it: you can check your specific language before committing a Workflow rather than discovering the answer in review.
**Best for** — content in languages outside the mainstream 30, and any workflow that needs non-speech events labeled rather than dropped.
**Strengths**
* **Speaker diarization up to 32 speakers** — the highest documented ceiling in the catalog, which matters for panels, roundtables, and large meetings.
* **Dynamic audio tagging** — non-speech events such as laughter and applause are captured as `audio_event` entries, which feeds accessibility caption requirements directly.
* **Entity detection across 65 types** and **keyterm prompting with up to 1,000 terms**, for enforcing domain vocabulary.
* Precise word-level timestamps, giving subtitle segmentation more to work with.
* A realtime variant transcribes 90+ languages at roughly 150 ms latency.
**Trade-offs**
* **Accuracy varies widely by language**, from under 5% WER to over 25%. Check ElevenLabs' published tier for your language rather than assuming the headline figure applies.
* On unclear or silent passages, transcription models of this class can produce fluent text that was never spoken. That failure mode does not look like an error in the transcript, so budget a validation pass for anything going straight to a deliverable.
**In Ollang** — `elevenlabs-scribe-v2` · Closed Captions, Subtitles, AI Dubbing.
***
## Speechmatics
A speech recognition specialist whose focus is accent and dialect coverage rather than benchmark position on clean, standard-accent English. That focus is why it holds up on audio that degrades generic models most.
Speechmatics covers **70+ individual languages**, plus **7 bilingual language packs** for media where two languages appear in the same file.
| Model | What it is for |
| ------------ | -------------------------------------------------------------------------------------------------------------------------- |
| **Enhanced** | Highest-accuracy operating point; supports translation |
| **Standard** | Faster operating point; supports translation |
| **Melia 1** | Multilingual model that switches between languages automatically, with no language selection; does not support translation |
**Best for** — batch transcription where speakers have strong regional accents, where lesser-resourced dialects appear, or where speakers move between languages spontaneously.
**Strengths**
* **Melia 1 handles spontaneous language switching without a declared source language** — the cleanest answer in the catalog for genuinely multilingual speakers.
* Consistently good on accented and dialectal speech, which is where generic models degrade most.
* Speaker diarization and custom dictionary for domain vocabulary.
* Smart formatting applied by default — numbers, dates, currencies, and measurements converted to conventional written forms — plus profanity tagging, disfluency removal, and regex-based word replacement.
* Translation from English into 34 target languages on the Enhanced and Standard models.
**Trade-offs**
* Little advantage over AssemblyAI or Scribe v2 on clean, standard-accent audio — its strengths only show up on the material that defeats generic models.
* Melia 1's automatic language switching and translation are mutually exclusive — pick the model that matches the job.
* 70+ languages is narrower than AssemblyAI's 99 or Scribe's 90+.
**In Ollang** — `speechmatics` · Closed Captions, Subtitles, AI Dubbing.
***
## Speechmatics with LLM
Speechmatics transcription with an LLM post-processing pass that accepts **custom guidelines and instructions**.
Everything true of [Speechmatics](#speechmatics) applies here, plus the ability to steer the output — enforce terminology, apply formatting conventions, or follow style guidance — instead of correcting it downstream. Speechmatics' own formatting controls are rule-based; this variant adds instruction-driven judgment on top.
**Best for** — recurring content with a house style: consistent speaker labels, fixed product terminology, specific number and date formatting, or client-mandated caption conventions.
**Trade-offs**
* Adds a post-processing stage, so it only earns its place when you would otherwise be making the same corrections by hand every week.
**In Ollang** — `speechmatics-llm` · Closed Captions, Subtitles, AI Dubbing.
***
## AWS Transcribe
Amazon's managed ASR service, strongest as an option for organizations already standardized on AWS.
Amazon Transcribe covers **more than 100 languages and locale variants** across batch and streaming, though feature support varies sharply by language.
**Best for** — enterprise pipelines that need transcription to sit inside an existing AWS footprint, with English variants and Spanish as the primary working languages.
**Strengths**
* **Nine English locales** — US, British, Australian, Indian, Irish, New Zealand, Scottish, South African, and Welsh — plus three Spanish locales, so regional variants are recognized rather than approximated.
* Batch and streaming modes with automatic language identification, including audio containing multiple languages.
* Custom vocabulary and acronym handling across most supported languages.
* PII redaction and Call Analytics for regulated material.
* AWS-native integration — IAM, VPC, KMS, and regional controls apply without additional plumbing.
**Trade-offs**
* **No general accuracy advantage** over the other single-model providers on ordinary media content — its case rests on AWS alignment and locale depth.
* **Custom language models cover only seven locales**: en-AU, en-GB, en-US, es-US, de-DE, ja-JP, and hi-IN (batch only). The customization story is narrower than it first appears.
* **Redaction is limited to English, French, Italian, Portuguese, and Spanish.**
* Streaming is unavailable for some languages and in some AWS regions.
**In Ollang** — `aws` · Closed Captions, Subtitles, AI Dubbing.
***
## Whisper
Whisper is **not currently selectable** in Workflows. It remains documented because it is the reference open-source baseline that most comparisons in this catalog are measured against, and because self-hosted Whisper deployments are common in customer-side pipelines.
OpenAI's open-source ASR model, trained on 680,000 hours of multilingual audio. It handles noisy, multilingual, and poor-quality recordings well for an openly available model, and it can be self-hosted, which matters for data-residency-constrained workloads.
Its well-documented weakness is hallucination on silence and low-quality segments — it will confidently emit text for audio that contains none. That failure mode is the reason the hosted providers above are preferred for deliverable-grade output.
**In Ollang** — `whisper` · Not currently available for selection.
***
## Related
Pipeline steps, provider tags, and status definitions.
How to configure these providers at global and folder level.