Ollang Agentic STT

Ollang Agentic STT is Ollang’s own transcription workflow and the default provider for every transcribe step. Unlike a single-model ASR service, it runs a collaborative multi-agent pass: several interpretations of the same audio are produced and reconciled, speaker attribution is resolved across the whole timeline, and surrounding context — including non-speech audio cues — is used to choose between readings that a single model would have to guess at.

​

Models

​

Key capabilities

• Multi-agent reconciliation — competing transcription hypotheses are compared rather than accepted, which is what produces the accuracy gain on ambiguous audio.
• Noise and accent robustness — built for real-world source material: field recordings, conference audio, overlapping speakers, and wide accent diversity.
• Multi-language streams — handles video and audio where more than one language is present, without forcing a single language assumption across the file.
• Contextual audio cues — non-speech information is treated as signal, which matters when the transcript becomes a dubbing script rather than a reading transcript.
• Stable speaker attribution — speaker identity holds across long-form content instead of drifting between segments.

​

Visual STT

Visual STT is a different modality in the same family. Instead of transcribing speech, it reads on-screen text from video frames and converts it into editable, timecoded text content. It is the right provider when the message lives in the picture rather than the audio:

• silent video and sound-off social content,
• burned-in captions and subtitles,
• title cards, lower thirds, and kinetic typography,
• product and UI demos where the interface text carries the meaning.

Because it transcribes what is displayed rather than what is said, content with meaningful speech and meaningful on-screen text usually needs both an audio pass and a visual pass.

​

Where it fits in a workflow

Agentic STT is the default because transcription errors do the most damage of any in the pipeline — everything downstream inherits them. A mistranscribed name becomes a mistranslated name and then a mispronounced one in the dub. Choose it when the source audio is genuinely difficult. On clean, single-speaker studio narration, AssemblyAI or ElevenLabs Scribe v2 deliver comparable transcripts in a single pass.

​

Trade-offs

• Throughput — a multi-agent pass does more work than a single model call, so it is slower on large batches. What it buys is robustness, not speed.
• Experimental tiers — Agentic STT-1.2 and Visual STT may change between releases. Run them in a benchmarking Folder alongside your production provider rather than switching a Global Workflow to them.

​

Reference

• Speech-to-Text provider catalog
• Workflows and provider architecture
• Order types