LLM Proxy

Appearance

Create transcription ​

Speech-to-text on an uploaded audio file. Multipart form data.

POST/v1/audio/transcriptions

multipart/form-data. Missing model/file or a non-numeric temperature → 400 with an OpenAI-shaped error.

Request
bash
curl http://localhost:8080/v1/audio/transcriptions \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -F file=@recording.mp3 \
  -F model=whisper-model \
  -F response_format=verbose_json \
  -F timestamp_granularities[]=word
Request — OpenAI SDK
python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-your-proxy-key")
text = client.audio.transcriptions.create(
    model="whisper-model",
    file=open("recording.mp3", "rb"),
)
print(text.text)
Response
json
{
  "task": "transcribe",
  "language": "en",
  "duration": 12.5,
  "text": "Hello from the proxy!",
  "segments": [
    {
      "id": 0,
      "start": 0.0,
      "end": 2.1,
      "text": " Hello from the proxy!",
      "words": [
        {"word": " Hello", "start": 0.0, "end": 0.4},
        {"word": " proxy!", "start": 1.7, "end": 2.1}
      ]
    }
  ]
}

Request fields ​

file (required), model (required), language, prompt, response_format (default json), temperature, stream, multi-value timestamp_granularities[] and include[]. Other form fields are forwarded.

Response formats ​

  • json → {"text": …} (default)
  • verbose_json → adds task, language, duration, segments, words
  • text / srt / vtt → text/plain
  • diarized_json → duration, segments, task

usage is included when reported (tokens or audio duration). stream: true returns SSE; the proxy adds include[]=usage upstream automatically.

Provider support ​

openai, the openai-compatible family (except qwen/qwen-intl, whose OpenAI-compatible ASR surface is /chat/completions with input_audio instead), gemini (native STT), and openrouter (STT) — see the capability matrix.

Request
bash
curl http://localhost:8080/v1/audio/transcriptions \
  -H "Authorization: Bearer $LLM_PROXY_KEY" \
  -F file=@recording.mp3 \
  -F model=whisper-model \
  -F response_format=verbose_json \
  -F timestamp_granularities[]=word
Request — OpenAI SDK
python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-your-proxy-key")
text = client.audio.transcriptions.create(
    model="whisper-model",
    file=open("recording.mp3", "rb"),
)
print(text.text)
Response
json
{
  "task": "transcribe",
  "language": "en",
  "duration": 12.5,
  "text": "Hello from the proxy!",
  "segments": [
    {
      "id": 0,
      "start": 0.0,
      "end": 2.1,
      "text": " Hello from the proxy!",
      "words": [
        {"word": " Hello", "start": 0.0, "end": 0.4},
        {"word": " proxy!", "start": 1.7, "end": 2.1}
      ]
    }
  ]
}

MIT licensed

LLM Proxy