# Whisper-1 (OpenAI API) - Core.Today AI API

> OpenAI's hosted whisper-1 speech recognition with the OpenAI API parameters - json or verbose_json output, segment and word timestamps (timestamp_granularities), language hint and prompt. Billed per second of audio (0.2325 credits/s, about 14 credits per minute), up to 1 hour / 25MB per request.

- **Provider**: OpenAI
- **Model ID**: openai/whisper-1
- **Category**: Audio & TTS
- **Credits**: 14 per minute of audio (0.2325/s, settled on measured length)
- **Speed**: Fast
- **Quality**: High

## Features
- OpenAI's hosted whisper-1 with the OpenAI API parameter names
- verbose_json output with language, duration and timestamped segments
- Segment and word timestamps via timestamp_granularities
- ~100 languages with auto detection, or an ISO-639-1 language hint
- Accepts flac, m4a, mp3, mp4, mpeg, mpga, ogg, wav and webm up to 25MB

## Use Cases
- Subtitles and captions from segment start/end timestamps
- Meeting, lecture and interview transcripts
- Migrating an existing OpenAI whisper-1 integration onto Core.Today credits

## API Endpoint

Base URL: https://api.core.today/v1
Create Prediction: POST /predictions
Get Status: GET /predictions/{job_id}

## Authentication
Header: X-API-Key: YOUR_API_KEY

## Input Parameters

### Required
- **audio**: string - Audio file URL (flac, m4a, mp3, mp4, mpeg, mpga, ogg, wav, webm). Sent to OpenAI as the `file` part. Max 25MB and 1 hour.

### Optional
- **response_format**: string (default: json) - json returns {text, usage}; verbose_json adds language, duration and timestamped segments/words. text, srt and vtt are not available through this API.
  Options: json, verbose_json
- **timestamp_granularities**: array - Timestamp detail for verbose_json - segment, word or both. Requires response_format verbose_json. Word timestamps add latency; segment timestamps add none.
  Options: segment, word
- **language**: string - Language of the input audio as an ISO-639-1 code (e.g. ko, en, ja). Improves accuracy and latency. Omit for automatic detection.
- **prompt**: string - Optional text to guide style or continue a previous segment. Should match the audio language.
- **temperature**: number (default: 0) - Sampling temperature between 0 and 1. 0 lets the model raise it automatically until certain thresholds are hit.

## Examples

### Segment timestamps (verbose_json)
Returns text, language, duration and segments with start/end seconds.
```json
{
  "model": "openai/whisper-1",
  "input": {
    "audio": "https://replicate.delivery/mgxm/e5159b1b-508a-4be4-b892-e1eb47850bdc/OSR_uk_000_0050_8k.wav",
    "response_format": "verbose_json",
    "timestamp_granularities": [
      "segment"
    ],
    "language": "en"
  }
}
```

### Plain transcript (json)
Returns {text, usage} for a Korean recording.
```json
{
  "model": "openai/whisper-1",
  "input": {
    "audio": "https://example.com/meeting.m4a",
    "response_format": "json",
    "language": "ko"
  }
}
```

## Response Format
```json
{
  "job_id": "abc123",
  "status": "pending | processing | completed | failed",
  "result": "URL or data (when completed)"
}
```

## Usage Flow
1. POST /predictions with model and input -> receive job_id
2. GET /predictions/{job_id} -> poll until status is completed or failed
3. Result contains output URL(s)

## Tips
- Retiring: OpenAI shuts whisper-1 down on 2027-02-26 (the translations endpoint too). openai/gpt-transcribe is the replacement for plain transcripts and is cheaper; it has no timestamps, so keep timestamp workflows on openai/whisper (Replicate segments)
- The result is OpenAI's response body as-is: json gives text and usage.seconds (the billed length); verbose_json gives text, language, duration and segments/words
- timestamp_granularities works only with verbose_json; when omitted, verbose_json returns segment timestamps
- Split files over 25MB or 1 hour before submitting - longer or larger files are rejected
- For SRT/VTT subtitles, build them from verbose_json segments, or use openai/whisper (Replicate) which outputs srt/vtt directly
- OpenAI SDK: point the client at base_url https://api.core.today/llm/openai/v1 with your cdt_ API key and call client.audio.transcriptions.create(model="whisper-1", file=..., response_format="verbose_json", timestamp_granularities=["segment"]) - the file is uploaded directly (no URL), json/verbose_json only, billed per second at the LLM rate (0.1858 credits/s, about 11.15 per minute) on the length OpenAI reports

## Documentation
https://platform.openai.com/docs/guides/speech-to-text