Accuracy expectations & service standards

What we promise on transcript quality, how we measure it, what we do automatically when a result comes back below its expected band, and what we'll do if something genuinely goes wrong on our end.

Last updated 2026-05-27 · Applies to all transcriptions processed by VTS

1. Accuracy bands

These are ranges we observe across real audio — not contractual guarantees. Audio quality, accents, overlapping speakers, and music all bound the ceiling. We auto-retry with stronger settings to push every transcript into its expected band, but no model produces a reliable transcript on heavy music or fully-degraded source audio.

Speech audio

Songs & music

No guaranteed accuracy means we'll still transcribe — and run the full escalation ladder below — but the result reflects what the source audio actually contains. For heavy music, distorted live-concert audio, or recordings dominated by background noise, no current model produces a reliable transcript at any price point.

2. How accuracy is measured

The percentage shown on each transcript is an estimated accuracy, not a measured word-error-rate. We compute it from Whisper's per-segment avg_logprob values combined with the detected-language probability — both signals that strongly correlate with actual transcription quality on real audio.

It's directional:

• 90%+ — model is highly confident in every segment. In practice, near-verbatim accuracy on clean speech.
• 70-90% — usable as-is for notes, captions, or quotation with light proofread. Most music-on-vocals lands here.
• 50-70% — directionally correct but needs human cleanup before publication. Common for noisy field recordings.
• Below 50% — best-effort output. Read with a grain of salt; consider a higher-quality recording.

The confidence number drives the auto-escalation ladder below: if a tier returns below its floor, the next tier picks up automatically.

3. What affects accuracy

In rough order of impact:

• Audio quality — bitrate, signal-to-noise ratio, dynamic range. A clean 64+ kbps voice recording out-performs a 32 kbps phone recording across every band.
• Background noise — fans, traffic, crowd hum, keyboard clatter, instrumental tracks. The denoise pre-pass in Attempt 2 handles most of this, but cannot recover speech that's drowned out at the source.
• Accents & dialects — Whisper handles most major accents reliably; very heavy regional dialects and code-switching can drop quality by 10-20 percentage points.
• Overlapping speakers — Whisper transcribes the dominant speaker; simultaneous speech is often dropped or merged. For interviews / panels, recordings with separate mics per speaker transcribe ~15 pts higher than a single room mic.
• Music presence — instrumentals mixed under vocals are the most disruptive single factor. Songs with melodic repetition (chorus, bridge) trigger the model's anti-hallucination guardrails. The denoise pre-pass + song-mode params on Attempt 2 push accuracy back up but cannot make a heavy-mix EDM track transcribable.
• Language switching mid-clip — Whisper detects language at the start and sticks with it. A clip that opens in English and switches to Spanish will lose the Spanish portion.

4. Auto-escalation ladder

Every transcription runs up to three tiers transparently. If a tier returns below its floor, the next tier picks up automatically — same transcript row, no extra charge, the user sees one transcript that progressively improves. Cap is three attempts so we don't loop forever on genuinely no-guarantee content.

During escalation, the transcript row's status reads "Auto-improving transcription — Attempt 2 of 3…" so you can see the system working. Final attempt's confidence is what shows on the transcript card.

5. What we transcribe

6. Language support

VTS transcribes any language Whisper supports — about 100 languages including English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Japanese, Korean, Chinese, Hindi, Arabic, Turkish, Vietnamese, Indonesian, Filipino, and many more.

Accuracy is highest on the languages Whisper was trained on most heavily (English, Spanish, French, German). Lower-resource languages may sit ~10-20 percentage points below the band ceilings even on clean audio. Language is auto-detected at the start of the clip and does not switch mid-stream.

7. Service availability

VTS is a best-effort service. We target high availability but do not publish a formal uptime SLA.

• Target uptime — 99% measured monthly, excluding scheduled maintenance and upstream provider outages (Stripe, YouTube, Facebook, HuggingFace, OpenAI).
• Scheduled maintenance — usually announced at least 24 hours ahead via a banner on the homepage. Most deploys are zero-downtime and don't require maintenance windows.
• Incident communication — for outages affecting transcription, we email signed-in users with active credit balances or active subscriptions once the incident is identified.
• Data durability — completed transcripts are stored in our managed Postgres with automated daily backups. We do not guarantee permanent retention; export your transcripts (.txt / .srt) if you need long-term archival.

8. Refund policy

Three tracks:

• PAYG → wallet credit. Failed transcriptions paid from your wallet have the exact cents returned to your wallet balance within a few minutes. The credit is idempotent (a retry can't double-credit) and the original charge appears in your transaction history alongside the reversal so the audit trail is intact.
• Pro subscription → minute credit. A failed transcription for a Pro user doesn't count against the 25-hour included pool — the minutes that were consumed by the failed run are added back to the current period's included quota. The transcription effectively "didn't happen" against your monthly pool. (If you were already past the soft cap and the failed run fell through to wallet PAYG, the credit goes back to the wallet instead — same as the PAYG track.)
• Quality issues → re-transcribe at no charge. If a completed transcript wasn't accurate enough, the auto-escalation ladder usually fixes it. If not, click Re-transcribe + denoise on the transcript row to force-run the higher tiers (denoise + medium / large model). No new charge, no new transcript row — the existing one is refreshed in place.

What we do not refund:

• Subscription charges to your card. Pro is a flat $15/month plan; we do not prorate, refund, or partially reverse subscription charges via Stripe. Cancel anytime in the Customer Portal and you keep Pro through the end of the current billing period; no further charges. Locked policy as of 2026-05-27.
• Successfully-completed transcripts you're unhappy with for stylistic reasons. The auto-ladder + the Re-transcribe button cover quality issues; we don't reverse charges for "I didn't like the formatting" or "the speakers weren't labeled how I expected".

Why no Stripe-side refunds? Money flowing back through Stripe creates accounting friction (Stripe fees aren't refunded, dispute risk goes up, your statement gets messy). In-system credits — wallet cents for PAYG, included minutes for Pro — are faster, reversible, and don't cost you anything in fees or cross-currency conversion. If Stripe themselves issue a refund (rare: a chargeback or a duplicate-charge dispute that you win), our books simply reflect the reversed charge: the corresponding wallet balance is debited to keep the audit trail straight.

9. Improving your audio

The single biggest accuracy lever is the audio you upload. Two minutes of attention upstream is worth more than any number of re-transcribes:

• Record close to the source. A lavalier mic six inches from the speaker beats a phone across the room by 20+ percentage points.
• Mute the music. If you're recording a podcast / interview / vlog with a soundtrack, record the speech track separately and add music in post. The denoise pre-pass helps but it's not a miracle.
• One mic per speaker for panels and multi-person interviews. Single-mic recordings of group conversations lose ~15 pts of accuracy versus per-speaker mics.
• Pick a quiet room. Tile bathrooms, parking garages, and reflective spaces add room echo that no denoiser fully removes. Soft furnishings + carpet = better transcripts.
• Use 16 kHz mono or higher. Whisper resamples to 16 kHz internally; under-sampled or lossy inputs lose detail that can't be recovered.
• Keep clips under 20 minutes when possible. Shorter clips trigger less hallucination cascade behaviour. For long content, splitting at natural breaks (chapters, scene changes) gives consistently higher accuracy than a single 2-hour file.

10. Reporting an issue

If a transcript came back wrong in a way the auto-escalation ladder didn't catch, or you think you were charged for something we shouldn't have charged for, email support@askgiya.com with the transcript ID (visible at the top of the transcript card) and a short description. We read every message.

For all other issues, the FAQ at /support is the fastest first stop.

This document is the source of truth for VTS accuracy expectations. The bands and the ladder are implemented in code at quality.py; this page renders from the same module so the numbers can't drift.