Add transcription to your own API

FunASR Blog · 2026-06-18 · Updated 2026-09-09

If your application needs to upload a recording and receive its text, start with that single round trip. A local transcription endpoint lets your team check the request and response before adding real traffic or a public gateway.

The example uses an authorized audio.wav: a client uploads the file to a local SenseVoice service, then reads the JSON text field. It supplies a callable path, not a fabricated transcript or a speed comparison.

This article covers the packaged funasr-server in FunASR 1.4.15. It offers part of the OpenAI transcription interface, not every hosted feature or response format; a working URL alone does not complete a migration.

Start with one local JSON transcription

python -c 'from importlib.metadata import version; assert version("funasr") == "1.4.15", version("funasr")'
python -m pip check

funasr-server --host 127.0.0.1 --port 8000 --model sensevoice --device cpu

From another terminal on the same host, upload an audio.wav file you are authorized to process. The command prints transcript content; keep private audio results out of public logs.

curl --fail --show-error --silent --max-time 120 \
  http://127.0.0.1:8000/v1/audio/transcriptions \
  -F "file=@audio.wav" \
  -F "model=sensevoice" \
  -F "response_format=json"

The response is a JSON object containing text. Its content depends on the recording and model. The 120-second client timeout is an example, not a duration guarantee or proof that timed-out inference is cancelled.

OpenAI Python SDK: specify the model and format

The following basic call uses an environment with the separate openai client installed. A migration must still check model names, fields, response parsing, errors and gateway authentication, not just URL reachability.

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="not-needed",
    timeout=120.0,
    max_retries=0,
)
with open("audio.wav", "rb") as audio:
    result = client.audio.transcriptions.create(
        model="sensevoice", file=audio, response_format="json"
    )
print(result.text)

The basic JSON call above does not produce subtitles. The packaged handler does not generate SRT/VTT; its text format returns a JSON string rather than plain text. Check the exact format contract in the appendix before changing the application parser.

Before connecting Open WebUI or another application

• Inspect the application's actual model alias, file type and response_format, plus how it handles success, empty text, errors and timeouts.
• A container's 127.0.0.1 is the container itself. A host loopback listener does not become reachable merely by changing a URL to host.docker.internal. Cross-container setups need a restricted private network and matching gateway authentication.
• Basic, Bearer and OIDC credentials are not interchangeable. Do not reuse the local placeholder as a production credential. Preloading one model is not a model-admission policy.
• Verify the entire path with controlled audio before admitting real traffic. Follow the maintained client recipes.

Compare complete deployments, not just API prices

Self-hosting has compute, storage, bandwidth, maintenance and monitoring costs. Data control depends on the client, proxy, temporary files, logs and retention policy; neither no-disk handling nor leak-free transcripts are automatic. Review model-weight licenses separately from the toolkit's code license.

Choose by language, latency, hardware and speaker requirements using model selection, MOSS transcription and diarization and the deployment matrix. This article supplies no new accuracy or performance ranking. For acceptance criteria, see the meeting-audio checklist.

Before another user connects, work through the service security guide and verify that the gateway cannot be bypassed. Keep the local example private until those deployment checks pass.