Using Deepgram Conversational STT / Flux (Python SDK)

Turn-aware streaming STT at /v2/listen — optimized for conversational audio (end-of-turn detection, eager EOT, barge-in scenarios).

When to use this product

• You're building a conversational UI and need explicit turn boundaries.
• You want Flux models (optimized for human-to-human or human-to-agent conversation).
• You want lower latency turn signals than v1 utterance_end.

Use a different skill when:

• You want general-purpose transcription (captions, batch, non-conversational) → deepgram-python-speech-to-text.
• You want a full interactive agent (STT + LLM + TTS) → deepgram-python-voice-agent.
• You want analytics (summarize/sentiment) → deepgram-python-audio-intelligence.

Authentication

import os
from dotenv import load_dotenv
load_dotenv()

from deepgram import DeepgramClient
client = DeepgramClient(api_key=os.environ["DEEPGRAM_API_KEY"])

Header: Authorization: Token <api_key>. WSS only — no REST path on v2.

Quick start

import threading, time
from pathlib import Path
from deepgram.core.events import EventType
from deepgram.listen.v2.types import (
    ListenV2CloseStream,
    ListenV2Connected,
    ListenV2FatalError,
    ListenV2TurnInfo,
)

with client.listen.v2.connect(
    model="flux-general-en",
    encoding="linear16",
    sample_rate="16000",
) as conn:

    def on_message(m):
        if isinstance(m, ListenV2TurnInfo):
            print(f"turn {m.turn_index} [{m.event}] {m.transcript}")
        elif isinstance(m, dict):                     # untyped fallback
            if m.get("type") == "TurnInfo":
                print(f"turn {m.get('turn_index')} [{m.get('event')}] {m.get('transcript')}")
        else:
            print(f"event: {getattr(m, 'type', type(m).__name__)}")

    conn.on(EventType.OPEN,    lambda _: print("open"))
    conn.on(EventType.MESSAGE, on_message)
    conn.on(EventType.CLOSE,   lambda _: print("close"))
    conn.on(EventType.ERROR,   lambda e: print(f"err: {type(e).__name__}: {e}"))

    def send_audio():
        for chunk in mic_chunks():                     # 80ms recommended
            conn.send_media(chunk)
            time.sleep(0.01)
        conn.send_close_stream(ListenV2CloseStream(type="CloseStream"))

    threading.Thread(target=send_audio, daemon=True).start()
    conn.start_listening()

Key parameters

No language parameter on v2 — language is implied by model (flux-general-en) or hinted via language_hint on multi.

For application-controlled turns, use eot_threshold="1.0" with a sufficiently large eot_timeout_ms, then call conn.send_force_end_turn() for the active turn. ForceEndTurn requires deployment enablement; see examples/16-transcription-force-end-turn.py.

Events (server → client)

• ListenV2Connected — connection established
• ListenV2ConfigureSuccess / ListenV2ConfigureFailure — mid-session config changes
• ListenV2TurnInfo — per-turn transcript + event (Update, EndOfTurn, EagerEndOfTurn, ...) + turn_index
• ListenV2FatalError — terminal error

Client messages: ListenV2Media, ListenV2Configure, ListenV2ForceEndTurn, ListenV2CloseStream.

Async equivalent

from deepgram import AsyncDeepgramClient
client = AsyncDeepgramClient()

async with client.listen.v2.connect(model="flux-general-en", ...) as conn:
    # same .on(...) handlers, then:
    await conn.start_listening()

API reference (layered)

1. In-repo reference: reference.md — "Listen V2 Connect".
2. AsyncAPI (WSS): https://developers.deepgram.com/asyncapi.yaml
3. Context7: library ID /llmstxt/developers_deepgram_llms_txt.
4. Product docs:
   • https://developers.deepgram.com/reference/speech-to-text/listen-flux
   • https://developers.deepgram.com/docs/flux/quickstart
   • https://developers.deepgram.com/docs/flux/language-prompting

Gotchas

1. /v2/listen, not /v1/listen. Different route, different client path (listen.v2 vs listen.v1).
2. Flux models only. nova-3, base, etc. will be rejected. Use flux-general-en or flux-general-multi.
3. No language parameter. Language is set by model choice. Use language_hint on flux-general-multi.
4. sample_rate is a STRING in the SDK (e.g. "16000").
5. Send ~80ms audio chunks for best turn-detection latency.
6. Close with send_close_stream(ListenV2CloseStream(type="CloseStream")) — not send_finalize (that's v1).
7. Messages may arrive as typed objects OR raw dicts — the SDK uses a tagged union with construct_type for unknowns. Handle both branches (see socket_client.py patch in .fernignore).
8. socket_client.py is patched / frozen (see .fernignore → src/deepgram/listen/v2/socket_client.py). Don't overwrite that manual patch during regeneration; treat other listen/v2 files as generated unless the regen workflow says otherwise.
9. Omit encoding/sample_rate for containerized audio (WAV, OGG, etc.) — the server detects them from the container.

Example files in this repo

• examples/14-transcription-live-websocket-v2.py
• tests/manual/listen/v2/connect/main.py

Related skills

• deepgram-python-speech-to-text — v1 general-purpose STT (REST + WSS)
• deepgram-python-voice-agent — full interactive assistant

Central product skills

For cross-language Deepgram product knowledge — the consolidated API reference, documentation finder, focused runnable recipes, third-party integration examples, and MCP setup — install the central skills:

npx skills add deepgram/skills

This SDK ships language-idiomatic code skills; deepgram/skills ships cross-language product knowledge (see api, docs, recipes, examples, starters, setup-mcp).