Scribe v3 and v3 Live: Transcription API Costs and Setup

Use Scribe v3 for a finished recording. Its model ID is scribe_v3, and it uses ElevenLabs’ POST /v1/speech-to-text API. It supports more than 90 languages, speaker diarization for up to 32 speakers, word or character timestamps, audio event tags and up to 1,000 keyterms. The number-format and secondary-language parameters help with mixed-language files and output formatting.

Choose Scribe v3 Live when the application must receive text during the conversation. Its ID is scribe_v3_live; it uses a WebSocket at /v1/speech-to-text/realtime and returns partial updates followed by committed transcript segments. Live adds turn prediction, multilingual switching within a sentence, streaming speaker diarization and a keyterm library. This is speech recognition, separate from Eleven Realtime, which generates speech for voice agents.

Turn prediction uses conversational context as well as silence to decide when a speaker has finished. The default commit strategy remains manual; request turn_prediction explicitly to enable it. Current API documentation supports that option, while some SDK guides still do not expose every Live feature. If a needed setting is missing from your SDK, connect to the documented WebSocket directly.

Upload a file as multipart form data and set model_id to scribe_v3. This minimal request returns the transcription response as JSON. Add timestamps_granularity=word when aligning words to subtitles or media, and diarize=true when speaker labels matter. The API accepts audio and video files; its transcription guide currently allows files below 5 GB and up to 15 hours in standard single-channel mode.

The v3 API is a one-line model-ID change for many existing Scribe v2 integrations. Check the returned transcript against names, numbers and short turns before treating it as publish-ready. For terminology, keyterms accepts up to 1,000 entries; the current price table lists an additional $0.05 per audio hour for batch keyterm prompting. The API reference applies a 20-second minimum billable duration when a request supplies more than 100 keyterms.

Open the realtime WebSocket from a server, sending your API key in the xi-api-key handshake header. Select scribe_v3_live and an audio format that matches the bytes you send. The example below shows one 16 kHz PCM chunk after the connection is open; the application continues sending chunks and reads partial_transcript and committed_transcript events. Use the same connection for the audio stream and transcript events.

To use turn prediction, add commit_strategy=turn_prediction to the connection URL. For diarization, add diarize=true; realtime labels can be revised as the service merges speaker identities, so apply speakers_merged events to earlier words. Diarization cannot be combined with background-audio filtering. Live audio is billed for all streamed seconds, including pauses and silence.

The ElevenAPI pricing page currently lists $0.20 per audio hour for Scribe v3 and $0.22 for Scribe v3 Live. At those published rates, ten recorded hours cost $2.00 and ten streamed hours cost $2.20 before optional feature charges. One hundred hours cost $20 or $22. For Live, count all seconds sent, including pauses; for multichannel uploads, the API bills each channel for the full audio duration.

Subscription plans include different amounts of transcription time. API usage is billed in USD; include your plan fee and usage beyond included hours when estimating the invoice. Enterprise terms can set different rates.

The current price table lists batch keyterm prompting at an additional $0.05 per audio hour and transcript editing at $0.07 per audio hour. For Live, keyterm prompting and the keyterm library are each $0.08 per hour, and editing is $0.12. A 30-minute batch job costs $0.10 at the base rate or $0.125 with keyterm prompting. Editing has a ten-second minimum per edited segment; count only the features and audio your request uses.

ElevenLabs reports 27% fewer word errors for v3 than Scribe v2 across 102 FLEURS languages. For v3 Live, it reports 44% lower semantic word error rate than Scribe v2 Realtime on a 73-language FLEURS evaluation, and a 3.1% diarization error rate on the English subset of its 16-language real-world set. These are publisher evaluations. The company averaged results by language and notes that its figures are not directly comparable with pooled corpus scores from other tests.

Cekura’s live benchmark page read on November 11 reports 1.89% overall word error rate for Scribe v3 Live across 1,000 public Pipecat clips and eight licensed Ocular recordings from four phone calls. The separate Ocular result is 2.05%. Median time to first text is 685 ms from streaming start; final-text delay is 57 ms after speech ends. Its latency evaluation covers 206 streamed turns, with 205 timed and one failed request. These measures describe that dataset and finalization method; they do not measure batch transcription, diarization accuracy or general voice-agent response latency.

Sonetho published a Korean comparison on November 11 using seven generated speech clips and the same scripts with Scribe v2 and v3. Both versions transcribed clean Korean readings with no character errors in those clips, while brand names and spoken numbers produced different text; v3 sometimes rendered English names phonetically in Hangul. The test used generated voices, a small sample and no live variant, so it is useful as a reminder to test names and number formatting rather than as a real-call accuracy estimate.

For a workflow decision, use a representative recording with a checked reference. Compare word errors, speaker assignments, timestamp alignment, response time and accepted output. Run a separate streaming test for v3 Live: the point of that model is to deliver usable partial and committed text during a conversation, not to win a file-only score.