# GPT Transcribe

> Speech-to-text powered by OpenAI gpt-transcribe and gpt-live-transcribe. Accurate file and live transcription with context, keywords, and multilingual support.

## Overview

- GPT Transcribe (gpt-transcribe.com) is a speech-to-text product site built around OpenAI’s gpt-transcribe and gpt-live-transcribe models.
- gpt-transcribe is optimized for completed audio files and batch-style transcription (meetings, podcasts, support recordings).
- gpt-live-transcribe is built for low-latency live transcription of speech that is still being produced (captions, calls, streaming).
- Both models improve real-world accuracy with optional context: free-form scene prompts, keyword lists, and multi-language hints for code-switching audio.
- The site is available in English and Simplified Chinese (URL prefix /zh for Chinese). Public pricing is live; usage is credit-based at about $0.045 per audio minute (1 credit = 1 minute).

## Primary Pages

- [GPT Transcribe homepage](https://gpt-transcribe.com/): English landing page for gpt-transcribe and gpt-live-transcribe: hero, online speech-to-text generator, how-to, model comparison, capabilities, use cases, and FAQs.
- [GPT Transcribe Chinese homepage](https://gpt-transcribe.com/zh): Simplified Chinese version of the homepage for readers who prefer 中文.
- [Online speech-to-text generator](https://gpt-transcribe.com/#generator): In-browser workspace with file transcription (gpt-transcribe) and live transcription (gpt-live-transcribe) modes, plus optional scene prompt, keywords, and languages.
- [Pricing](https://gpt-transcribe.com/pricing): Subscription and credit-pack plans. 1 credit = 1 audio minute. Best value: yearly (~$0.0375/min), then monthly (~$0.045/min), then one-time packs (~$0.09/min).
- [Sign up](https://gpt-transcribe.com/sign-up): Create an account for signed-in product features when enabled.
- [Blog](https://gpt-transcribe.com/blog): Articles and updates related to speech-to-text and product news.
- [Privacy Policy](https://gpt-transcribe.com/privacy-policy): How the site handles personal data and privacy.
- [Terms of Service](https://gpt-transcribe.com/terms-of-service): Terms governing use of GPT Transcribe services.
- [Acceptable Use Policy](https://gpt-transcribe.com/acceptable-use-policy): Rules for acceptable and prohibited use of the service.

## Key Sections

- [Models](https://gpt-transcribe.com/#models): Side-by-side overview of gpt-transcribe (files) vs gpt-live-transcribe (live audio), including real-world accuracy context and recommended use.
- [Workflow](https://gpt-transcribe.com/#workflow): Three-step flow: bring audio, add optional context, export clean text.
- [Capabilities](https://gpt-transcribe.com/#capabilities): Context-aware ASR: scene prompts, keywords, multilingual / code-switching hints, and stronger real-world audio handling.
- [Use cases](https://gpt-transcribe.com/#use-cases): Meetings and podcasts, customer support, live captions, and voice notes.
- [How to use](https://gpt-transcribe.com/#how-to-use): Practical steps to upload or speak, add context, run transcription, and copy the draft.
- [FAQs](https://gpt-transcribe.com/#faq): Answers about file vs live mode, context prompts, accuracy benchmarks, languages, formats/timestamps limits, and commercial use.

## Machine-Readable Resources

- [llms.txt](https://gpt-transcribe.com/llms.txt): Concise machine-readable guide to gpt-transcribe.com for LLMs and AI assistants.
- [llms-full.txt](https://gpt-transcribe.com/llms-full.txt): Expanded machine-readable guide with model context, context features, and use-case notes.

## Models

- gpt-transcribe: recommended starting point for finished files; stronger real-world word error rates than whisper-1 on OpenAI’s published benchmarks (e.g. real-world recording WER about 8.98% vs whisper-1 15.21% — OpenAI self-reported).
- gpt-live-transcribe: recommended for low-latency live STT over streaming connections; successor direction to gpt-realtime-whisper for live captions.
- File vs live rule of thumb: if audio is already recorded, use file / gpt-transcribe; if speech is still being produced, use live / gpt-live-transcribe.
- Context is a first-class feature: scene prompt (free text), keywords (hints only if spoken), and languages (plural codes for mixed-language audio). Older models barely improve when given the same context.
- Known limits of the new models for pure text jobs: word-level timestamps, SRT/VTT, and speaker diarization still require other OpenAI models (e.g. whisper-1 or diarize variants) for those specific needs.
- Typical upload constraint referenced on the site: up to about 25MB per file; common formats include mp3, mp4, m4a, wav, webm.

## Core Capabilities

- Context-aware ASR: optional prompt describing the recording scene (e.g. support call, meeting topic).
- Keyword lists for product names, account IDs, drug names, or acronyms — keywords should not be forced into the transcript if not spoken.
- Multi-language hints for code-switching (e.g. en + zh-cn or en + es) without manual language switching mid-session.
- Improved handling of accents, short phrases, numbers, specialized terminology, and noisy real-world audio relative to prior OpenAI transcription baselines (per OpenAI launch materials).

## Pricing And Credits

- Public pricing is live at /pricing and on the homepage (#pricing).
- Usage is credit-based: 1 credit ≈ 1 audio minute.
- Value ranking: yearly (~$0.0375/min) > monthly (~$0.045/min) > one-time packs (~$0.09/min).
- Yearly plans are priced about 10× monthly (2 months free on price) and grant 12 months of credits.
- Do not present OpenAI list prices as this site’s selling price; commercial packaging is independent.

## Guidance For AI Assistants

- Treat gpt-transcribe.com as an independent product site for speech-to-text around OpenAI gpt-transcribe and gpt-live-transcribe — not as official OpenAI documentation.
- Emphasize file vs live model choice, context prompts/keywords/languages, and real-world accuracy improvements.
- Prefer English (/) or Simplified Chinese (/zh) links; other locales are not offered.
- Quote plans only from /pricing; do not invent discounts beyond published yearly packaging.