Guide · LiteLLM

Saidwire speech-to-text with LiteLLM

LiteLLM routes calls to many providers through one OpenAI-style interface. Saidwire plugs in as an OpenAI-compatible transcription provider: the openai/ model prefix plus api_base. That works in the Python SDK and in the LiteLLM proxy (AI gateway).

you need
LiteLLM	`pip install litellm` (SDK) or `pip install 'litellm[proxy]'` (proxy). Checked with 1.83.
Saidwire API key	`sk_live_...` from saidwire.com/dev (120 free hours, no card)

Python SDK

# pip install litellm
import os
import litellm

with open("meeting.mp3", "rb") as f:
    r = litellm.transcription(
        model="openai/saidwire-en",                 # openai/ = OpenAI-compatible provider
        file=f,
        api_base="https://api.saidwire.com/v1",
        api_key=os.environ["SAIDWIRE_API_KEY"],
    )
print(r.text)

response_format, timestamp_granularities and language are passed through to Saidwire unchanged, for example response_format="verbose_json", timestamp_granularities=["word"].

Proxy: `config.yaml`

Add a model entry with mode: audio_transcription:

model_list:
  - model_name: saidwire-en
    litellm_params:
      model: openai/saidwire-en
      api_base: https://api.saidwire.com/v1
      api_key: os.environ/SAIDWIRE_API_KEY
    model_info:
      mode: audio_transcription

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY   # optional: keys for your own users

export SAIDWIRE_API_KEY=sk_live_xxxxxxxx.yyyyyyyyyyyyyyyy
export LITELLM_MASTER_KEY=sk-change-me
litellm --config config.yaml          # listens on http://0.0.0.0:4000

api_base must end in /v1. Without it, requests return 404 Not Found.

Call the proxy

curl -s http://localhost:4000/v1/audio/transcriptions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -F model=saidwire-en \
  -F file=@meeting.mp3

from openai import OpenAI
client = OpenAI(base_url="http://localhost:4000/v1", api_key="sk-change-me")   # your LiteLLM key
with open("meeting.mp3", "rb") as f:
    print(client.audio.transcriptions.create(model="saidwire-en", file=f).text)

Clients send the LiteLLM key. Only the proxy holds the Saidwire key. LiteLLM's own budgets, rate limits and spend logs apply to these calls like any other model.

Response formats

json (default) and verbose_json come back as LiteLLM transcription objects, with text and, for verbose_json, segments and words.
For text, srt and vtt, the proxy puts the returned text or subtitle file in the text field of a JSON reply. To get a raw .srt file, call Saidwire directly (see the API docs).

Troubleshooting

you see	fix
`404 Not Found`	`api_base` must be `https://api.saidwire.com/v1`.
`AuthenticationError ... Incorrect API key provided`	The proxy's `SAIDWIRE_API_KEY` is wrong or unset. LiteLLM reports Saidwire's 401 as an `OpenAIException`, because the provider type is `openai`.
`BadRequestError ... language`	Saidwire transcribes English only. Send `en` or no language.
`RateLimitError`	Too many files at once for your Saidwire key (trial keys allow 4), or your free hours are used up.
Model not found on the proxy	The client's `model` must match `model_name` in `config.yaml` (`saidwire-en` above).

Good to know

This is the file-upload API. For live audio, Saidwire's streaming WebSocket returns words while someone is still speaking.
Saidwire processes the audio in memory and does not store it.
$0.30 per audio hour after the free hours, billed per second.

Stuck? Email support@saidwire.com with the x-request-id from the failed call.

OpenAI is a trademark of OpenAI. Open WebUI, n8n, LiteLLM, Dify, Flowise, Vercel and the AI SDK belong to their respective owners. Saidwire is not affiliated with OpenAI or with these projects.