DEV Community

1DS22CS119_Manish
1DS22CS119_Manish

Posted on

ShrutiScript Media: Fictional Transcriber Shortlist for 5 April 2027

ShrutiScript Media: Fictional Transcriber Shortlist for 5 April 2027

Fictional recruitment exercise. Every candidate, test clip, practical submission and score below is invented. This is not a real hiring decision. Rohit Kulkarni and Sneha Iyer: please review the rubric, language QA and proposed screening schedule. AssemblyAI confidence is used only to route transcript review—not to rank people.

Download the complete evaluation pack, source audio and verbatim transcripts

Ranked shortlist

The practical rubric allocates 30 points to language fidelity/completeness, 20 to speaker labels/timestamps, 20 to names/numbers/terminology, 15 to punctuation/formatting, and 15 to file hygiene/turnaround. Scores are simulated assessments of the fictional practical submissions, not derivations from ASR confidence.

Rank Candidate Language Score /100 Recommended next step
1 Rahul Verma Hindi 93 Fast-track to final panel
2 Aditi Rao English 92 Fast-track to final panel
3 Divya Raghavan Tamil 89 Final panel after native-language QA
4 Kunal Desai Gujarati 87 Final panel after native-language QA
5 Nikhil Jadhav Marathi 85 Final panel after native-language QA
6 Harpreet Kaur Punjabi 83 Calibration interview / second practical
7 Ishita Banerjee Bengali 81 Calibration interview / second practical
8 Anagha Bhat Kannada 79 Reserve; targeted practical retest
9 Meera Nair Malayalam 77 Reserve; targeted practical retest
10 Sai Charan Telugu 75 Reserve; targeted practical retest

Supplementary reserve: Sana Qureshi — Urdu, 82/100. This sample is not ranked while I confirm the exact ninth-Indian-language roster. The core set uses Punjabi; retaining the Urdu sample avoids discarding useful evidence if the roster changes.

What the AssemblyAI pass showed

All 11 clips completed transcription. English and Hindi are clean and complete. Several other outputs need native-language review because of truncation, wrong-script output or lexical substitutions. These are machine-output issues to investigate, not judgments about candidates or natural-speech performance.

Candidate / track AssemblyAI confidence Review required
Aditi Rao — English 0.9758 Complete English output; routine check of date/time formatting and organisation name.
Rahul Verma — Hindi 0.9950 Complete Hindi output; very high reported confidence; routine native review.
Nikhil Jadhav — Marathi 0.8444 Complete instruction sequence, but Marathi spelling, named entities and numeric rendering need review.
Ishita Banerjee — Bengali 0.7480 Review required: returned mainly Devanagari rather than Bengali script, year reads 2007, and ending is truncated.
Kunal Desai — Gujarati 0.8146 Review required: mainly Devanagari rather than Gujarati script; instruction sequence otherwise mostly present.
Divya Raghavan — Tamil 0.8694 Tamil script returned with substitutions; final verification/deadline instruction is truncated.
Sai Charan — Telugu 0.6848 Priority review: substantial truncation and mixed Kannada/Tamil/Latin output; low reported confidence.
Anagha Bhat — Kannada 0.7797 Priority review: output stops during “monsoon audit”; Kannada/Latin mix and lexical substitutions.
Meera Nair — Malayalam 0.6673 Priority review: mixed Tamil/Malayalam/Latin output and truncated ending; low reported confidence.
Sana Qureshi — Urdu 0.8073 Reserve track: Urdu script retained but multiple lexical substitutions; full human verification required.
Harpreet Kaur — Punjabi 0.8145 Review required: Devanagari rather than Gurmukhi and transcript ends during “verified transcript”; deadline missing.

The evaluation pack preserves every source script, raw transcript, transcript ID, duration, audio link and review flag. An alternate Universal-2 pass on eight weaker outputs confirmed the same main limitations; it did not silently replace the originals.

Reviewer decisions requested

Rohit Kulkarni

  1. Calibrate the 100-point rubric and decide whether ranks 1–5 should form the final interview panel.
  2. Confirm whether the primary screening remains Monday, 5 April 2027, with Tuesday, 6 April only as an accessibility/overflow option.
  3. Approve equivalent clip versions, time limits, reviewer rotation and decision deadlines for any second-date screening.

Sneha Iyer

  1. Assign native-language review for Bengali, Gujarati, Tamil, Telugu, Kannada, Malayalam and Punjabi outputs, especially truncation and script mismatches.
  2. Confirm the core nine-Indian-language roster and whether Urdu remains a reserve or replaces another track.
  3. Approve the style rules for dates, speaker labels, “monsoon audit”, native scripts, and Urdu right-to-left presentation.

Proposed screening flow

  • 09:30–10:00 IST: reviewer calibration; no candidate details discussed outside the authorised panel.
  • 10:00–12:30 IST: five final-panel conversations, 25 minutes each plus transitions.
  • 13:30–15:00 IST: targeted practical rechecks for the remaining five core candidates.
  • 15:15–16:00 IST: joint scoring, conflict-of-interest check and documented decision.

If a second date is required, use equivalent—not reused—audio, the same instructions and time limits, calibrated reviewers, and no detailed candidate feedback before all screening is complete. The accompanying Workplace Stack Exchange draft asks practitioners how to make that two-date arrangement fair.

Safeguards

  • Keep source scripts and raw ASR output immutable; corrections belong in a separate reviewed layer.
  • Do not infer candidate quality from synthetic voices, ASR confidence, accent or language.
  • Score only the practical evidence against the shared rubric, and record native-reviewer decisions.
  • Keep this fictional dataset out of real applicant tracking, training data and employment decisions.

Top comments (0)