Transcribe a meeting with speaker labels

One upload, one pass, and a transcript that attributes every line — with the timings, and with a read on which turns were close enough to a microphone to quote.

Before you start

An API key, kept server-side

Create one under Developer → API keys in the console. See Authentication.

A diarizing model this deployment serves

GET /v1/audio/models and take an id with supports_diarization: true. The examples use nur-stt-v1, which is that model.

A recording under about twenty minutes

The model handles roughly ninety minutes in one pass, but the production GPU sets the real ceiling nearer twenty. Split longer recordings at silence boundaries — see the last section.

The recipe

Speaker labels, word timings and the audibility read only come back in verbose_json. json, text, srt and vtt carry the words alone.

curl https://api.eesi.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $EESI_API_KEY" \
  -F file=@meeting.m4a \
  -F model=nur-stt-v1 \
  -F response_format=verbose_json \
  -F "timestamp_granularities[]=segment" \
  -F "timestamp_granularities[]=word"

from eesi_sdk import EESIClient

with EESIClient() as client, open("meeting.m4a", "rb") as source:
    result = client.create_transcription(
        file=("meeting.m4a", source.read(), "audio/mp4"),
        model="nur-stt-v1",
        response_format="verbose_json",
        timestamp_granularities=["segment", "word"],
    )

for segment in result["segments"]:
    if segment.get("audibility") == "faint":
        continue  # too quiet to trust — see below
    stamp = f"{int(segment['start']) // 60:02d}:{int(segment['start']) % 60:02d}"
    print(f"[{stamp}] {segment['speaker']}: {segment['text']}")

import { readFile } from "node:fs/promises";
import { EESIClient } from "@eesi/sdk";

type Segment = {
  start: number;
  end: number;
  text: string;
  speaker: string;
  audibility?: "on_mic" | "off_mic" | "faint";
};

const client = new EESIClient();

const result = (await client.createTranscription({
  file: new Blob([await readFile("meeting.m4a")], { type: "audio/mp4" }),
  model: "nur-stt-v1",
  responseFormat: "verbose_json",
  timestampGranularities: ["segment", "word"],
})) as { duration: number; segments: Segment[] };

for (const segment of result.segments) {
  if (segment.audibility === "faint") continue;
  console.log(`[${segment.start.toFixed(1)}] ${segment.speaker}: ${segment.text}`);
}

What comes back

{
  "task": "transcribe",
  "language": "en",
  "duration": 1284.5,
  "text": "Right, shall we start? ...",
  "noise_floor_db": -52.4,
  "speech_seconds": 1102.3,
  "silence_seconds": 182.2,
  "segments": [
    {
      "id": 0,
      "start": 0.0,
      "end": 2.7,
      "text": "Right, shall we start?",
      "speaker": "S01",
      "pace_wpm": 96,
      "energy_db": -18.2,
      "level_db": -21.5,
      "snr_db": 30.9,
      "audibility": "on_mic"
    }
  ],
  "words": [
    { "word": "Right", "start": 0.08, "end": 0.34, "segment_id": 0, "speaker": "S01" }
  ]
}

One pass produces all of it. There is no separate diarization stage and no forced aligner to run.

Field	Says
`speaker`	An anonymous label, numbered by order of first appearance in this file
`audibility`	`on_mic`, `off_mic`, or `faint` — how far the turn sat above the room’s noise floor
`snr_db` / `level_db`	The measurements behind that judgement, from the raw audio before normalization
`pace_wpm`	Words per minute over the turn
`energy_db`	Loudness measured on the audio the model heard, so it ranks delivery within the take
`gaps`	Silences between turns, with `speech_seconds` and `silence_seconds`

Do not quote a turn marked faint. The label means the speech barely rose above the room tone, which is exactly where a transcript is most likely to be wrong and most likely to be confidently wrong. off_mic is usable with care — someone across the table. on_mic is what you quote.

Word confidence, where it appears, is acoustic support at the aligned position — how strongly the audio backs a word that was already transcribed — not the recognizer’s probability that the word is right. The two differ measurably. If word timings could not be produced at all, the payload says so with words_unavailable: true rather than omitting the key, because an absent key is indistinguishable from an older backend.

Save the run, then name the speakers

The OpenAI-compatible route above is a pure relay: it returns the transcript and keeps nothing. POST /v1/speech/transcriptions does the same work and keeps it — one upload, because sending a ninety-minute file twice is the thing worth avoiding.

curl https://api.eesi.ai/v1/speech/transcriptions \
  -H "Authorization: Bearer $EESI_API_KEY" \
  -F file=@meeting.m4a \
  -F model=nur-stt-v1 \
  -F stream=false

That returns a session id. stream defaults to true, which streams diarized segments back as server-sent events as the file decodes — good for a progress bar, more work for a script, so this recipe turns it off. Now replace S01 and S02 with real names. Operator names are kept apart from the model’s own labels, so renaming never edits the transcript:

curl -X PATCH https://api.eesi.ai/v1/speech/transcriptions/$SESSION_ID/speakers \
  -H "Authorization: Bearer $EESI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"names": {"S01": "Dana Okafor", "S02": "Priya Raman"}}'

And the run is a library entry from then on:

You want	Call
Every saved run, newest first	`GET /v1/speech/transcriptions`
One run, with its full result and speaker names	`GET /v1/speech/transcriptions/{session_id}`
Play the source audio back	`GET /v1/speech/transcriptions/{session_id}/audio`
Erase it	`DELETE /v1/speech/transcriptions/{session_id}`

Playback returns a short-lived signed URL rather than redirecting to one, because a browser will not replay a bearer token onto a media request — an <audio src> needs the URL itself. It lasts an hour.

Where diarization gets it wrong

Speaker separation needs enough continuous speech per voice to be confident. Two failures are worth expecting rather than discovering:

One person split in two. On short fragments — a meeting of one-word interjections, a heavily cut recording — the same voice can land under two labels. More continuous speech per person fixes it; nothing in the request does.
Labels do not carry across files. Speakers are numbered by order of first appearance in this upload, so whoever talks first is S01 here and might be S02 in the next file. If you split a recording, do not assume S01 is the same person in part two — reconcile by listening, or by the names you set on each run.

Recordings longer than one pass

Split at silence boundaries, not at a fixed clock time: a cut mid-word costs you a word in both halves and can invent a speaker change at the seam. The response’s gaps array from a first short pass, or any silence detector, gives you the boundaries. The upload caps are 25 MB on an ordinary transcription model and 1 GB on a diarizing one, and a long-form request may decode for up to 30 minutes before it answers — set your client’s read timeout above that or the request is cut after you have paid for the pass. Over the cap you get 413 content_size_limit_exceeded, whose message names the alternative model if this deployment serves one.

What it costs

$0.001 per minute of audio, metered on the recording’s duration. A twenty-minute meeting is two credits, which rounds to two cents. See Pricing and limits. Saving a run stores the audio, which creates a retention duty that is yours rather than the platform’s — DELETE is a real erasure path. See Compliance.

Audio intelligence

Every field in the diarized response, in full.

Speech to text

The plain transcription path, and the language hint.

Errors

The 413, and what to do about it.

Compliance

Recording consent, and the deletion routes.

Build

Cookbook

Billing

Transcribe a meeting with speaker labels

Before you start

The recipe

What comes back

Save the run, then name the speakers

Where diarization gets it wrong

Recordings longer than one pass

What it costs

Next

Audio intelligence

Speech to text

Errors

Compliance

​Before you start

​The recipe

​What comes back

​Save the run, then name the speakers

​Where diarization gets it wrong

​Recordings longer than one pass

​What it costs

​Next

Audio intelligence

Speech to text

Errors

Compliance

Before you start

The recipe

What comes back

Save the run, then name the speakers

Where diarization gets it wrong

Recordings longer than one pass

What it costs

Next