Three-recognizer research system using confidence-assisted word voting. Studied for overlapping speech and explicitly not promoted to streaming production.
Authors explicitly state that no streaming harness was built to measure overlap-routing responsiveness.
Batch processing
Not collected
Fine-tuning
Not collected
Knowledge cutoff
Not collected
Release date
Not collected
Parameters
Not collected
Architecture
Confusion-network ROVER word voting across compact bilingual ASR, Greek-adapted Canary and an earlier LoRA Whisper large-v3 adaptationSource for Architecture (opens in a new tab)
Tie-breaking uses mean word confidence. This is a composite research system, not a new monolithic weight checkpoint.
General evaluation includes both; overlapping-speech experiments cover English only and are not extended to Greek.
Context caching
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Access & pricing
Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.
No public immutable checkpoint selector or standalone API contract verified for this dated/internal system.
Access
Research-only ensemble described in the paper, with no downloadable weights or public standalone API. Contact the authors for any evaluation availability.
Regions
Original experimental deployment location not specified
Endpoint pricing has not been collected. Check the provider for applicable rates.
Data & enterprise terms
The paper releases no model weights, data or internal tooling. No standalone processing agreement or data-retention contract is established for this paper-only/datetime-specific record; current preview policies are not backfilled.
Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.
01 / INITIAL INPUTProcess & retain
The system processes eligible input and retains reusable state.
02 / LATER REQUESTCheck available state
Reuse depends on matching input, session state and the cache policy.
03 / CACHE HITReuse & generate
Reusing state can reduce repeated work. Any charges follow the billing rules below.
KIEFER SA · paper-documented system
Not verified
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
How it works
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Minimum reusable input
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Lifetime & refresh
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
How to enable it
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Usage fields
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Billing rules
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
Limits & data boundaries
Not collected
No model-specific prompt-cache contract verified. Server-side audio segmentation, speech buffers and reused decoders do not establish discounted token caching.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
What the measurements say
ROVER · English overlapping-speech region
KIEFER SA / Sophea authors · provider-reported
37.87% WER
Paper reports reduction from 53.35% to 37.87% WER on overlap-only English audio. Not whole-meeting WER; Canary AMI training exposure is acknowledged and Greek overlap was not evaluated.
Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
The record behind the model
Why it is included
The provider-authored production paper identifies internal adapted speech models and combined systems and explicitly releases none of their weights, data or tooling. This record covers the specified proprietary adaptation/system; the original Qwen, Whisper and Canary base weights and stock-weight Turbo are not reclassified as closed.
