BILINGUAL RECOGNITION EVALUATION WORKSHEET Free resource. Reviewed October 20, 2026. No recording, paid call or benchmark run. This is a proposed evaluation. All result fields intentionally blank. 1. TASK AND RIGHTS Business task: ____ Languages and competent reviewers: ____ Permission to record/use each voice: ____ Source-word transcript required? ____ Separate translation required? ____ Original artifact storage/access/retention owner: ____ Do not use private calls or sensitive data merely to complete the example. 2. MATCHED CONFIGURATION Provider A/exact model: ____ Provider B/exact model: ____ API/adapter versions and dates: ____ Actual application controls supported: ____ Language bias settings: ____ Strict restriction enabled and why: ____ Audio codec/sample rate/channel count/chunking: ____ Same input-file hashes on both sides: ____ Options, context and normalization policy approved before scoring: ____ Country/region, run date and authorized budget: ____ 3. ORIGINAL FICTIONAL SCRIPT SPECIFICATION Record only after obtaining permission, with reviewers approving natural speech. A. Inside-clause switch: I need una visita por la tarde, but not on Friday. B. Corrected name: The name is Nila, not Mila. C. Negative instruction: Please do not arrange a visit; I only need the hours. D. Monolingual control: Can someone confirm the office opening hours? Vary actual speaker/accent/pauses only by a documented plan. These are script ideas, not spoken audio or measured provider outputs. 4. REFERENCE REVIEW Recording ID/hash: ____ Reviewer 1/reference: ____ Reviewer 2/reference: ____ Disagreement and approved resolution: ____ Mark original-language spans and exact switch boundaries: ____ Corrected name/negation/task-critical fields: ____ Permitted punctuation/number/contraction normalization: ____ Word/character/span scoring rule suitable for each language: ____ 5. RESULT ROW (COPY PER RECORDING AND PROVIDER) Model/settings: ____ Returned original-language words: ____ Switch span and error: ____ Negation preserved? ____ Name correction preserved? ____ Unintended translation? ____ Missing/failed transcript? ____ Provisional versus final artifact? ____ Reviewer rationale: ____ WER/CER/span score and denominator: ____ Source speech-end timestamp: ____ Final transcript receipt: ____ Latency definition/result: ____ Failed cases retained in report? ____ Provider units and billed amount: ____ Evidence file: ____ Accepted/failed/unknown: ____ Correction and retest scope: ____ 6. OPTIONAL TRANSLATION REVIEW, SEPARATE FROM RECOGNITION Approved target language and reviewer: ____ Link to exact source transcript/turn: ____ Translation preserves correction and exclusion? ____ Translation cost/latency measured separately? ____ Do not overwrite the original words with a translated staff note. 7. COMPARISON DECISION Matched recordings and settings? ____ Per-language/switch-span errors separated from global aggregate? ____ Missing transcripts and failures reported? ____ Uncertainty/sample size adequate for claimed difference? ____ Actual relevant caller channel tested? ____ Claim supported: ____ Claim still unknown: ____ Reviewer/date: ____ External general STT benchmarks do not fill these bilingual result fields. A recognition model alone does not prove a bilingual spoken reply or accepted action. Native application options may be narrower than direct provider APIs. Primary references, checked October 20, 2026: https://soniox.com/docs/stt/concepts/language-hints https://soniox.com/docs/stt/concepts/language-restrictions https://www.assemblyai.com/docs/streaming/multilingual-transcription Original guide: https://burki.dev/blog/soniox-vs-assemblyai-bilingual-transcription