Skip to main content
For real-time applications where time-to-first-byte (TTFB) is critical, use streaming mode. Streaming returns a sequence of server-sent events containing base64-encoded audio chunks, so playback can start before generation finishes. For the lowest possible interactive latency (and bidirectional text input), see the WebSocket API.

​
Streaming audio

from together import Together

client = Together()

# Save the streamed audio to a file
with client.audio.speech.with_streaming_response.create(
    model="canopylabs/orpheus-3b-0.1-ft",
    input="The quick brown fox jumps over the lazy dog",
    voice="tara",
    stream=True,
    response_format="raw",  # Required for streaming
    response_encoding="pcm_s16le",  # 16-bit PCM for clean audio
) as response:
    response.stream_to_file("speech_streaming.pcm")

​
Streaming response format

When stream: true, the API returns a stream of server-sent events. Audio chunk: 
data: {"type":"conversation.item.audio_output.delta","item_id":"tts_1","delta":"<base64_encoded_audio>"}
Word timestamps (when alignment=word): 
data: {"type":"conversation.item.word_timestamps","words":["Hello","world"],"start_seconds":[0.0,0.4],"end_seconds":[0.4,0.8]}
Stream end: 
data: [DONE]
When streaming is enabled, only raw (PCM) format is supported. For non-streaming requests, you can use mp3, wav, or raw.

​
Output raw bytes

If you want to extract raw audio bytes (for example, to feed into a custom audio pipeline), use the settings below. 
import requests
import os

url = "https://api.together.ai/v1/audio/speech"
api_key = os.environ.get("TOGETHER_API_KEY")

headers = {"Authorization": f"Bearer {api_key}"}

data = {
    "input": "This is a test of raw PCM audio output.",
    "voice": "tara",
    "response_format": "raw",
    "response_encoding": "pcm_s16le",
    "sample_rate": 24000,
    "stream": False,
    "model": "canopylabs/orpheus-3b-0.1-ft",
}

response = requests.post(url, headers=headers, json=data)

with open("output_raw.pcm", "wb") as f:
    f.write(response.content)

print(f"Raw PCM audio saved to output_raw.pcm")
print(f"   Size: {len(response.content)} bytes")
This writes the raw bytes to a test2.pcm file.

​
See also