Streaming

1
# ci:skip — requires sounddevice + an audio device, not available in CI
2
# Plays audio chunks as they arrive AND saves the full stream to streamed.wav.
3
# Requires: pip install websockets sounddevice numpy
4
import asyncio
5
import base64
6
import json
7
import os
8
import wave
9
10
import numpy as np
11
import sounddevice as sd
12
import websockets
13
14
API_KEY = os.environ["SMALLEST_API_KEY"]
15
WS_URL = "wss://api.smallest.ai/waves/v1/tts/live"
16
SAMPLE_RATE = 24000
17
18
async def stream_tts(text):
19
    audio_chunks = []
20
    # Open a non-blocking output stream so each chunk plays as it arrives.
21
    stream = sd.OutputStream(samplerate=SAMPLE_RATE, channels=1, dtype="int16")
22
    try:
23
        stream.start()
24
        async with websockets.connect(
25
            WS_URL,
26
            additional_headers={"Authorization": f"Bearer {API_KEY}"},
27
        ) as ws:
28
            await ws.send(json.dumps({
29
                "text": text,
30
                "voice_id": "meher",
31
                "model": "lightning_v3.1_pro",
32
                "sample_rate": SAMPLE_RATE,
33
            }))
34
35
            while True:
36
                data = json.loads(await ws.recv())
37
                status = data.get("status")
38
                if status == "chunk":
39
                    pcm = base64.b64decode(data["data"]["audio"])
40
                    audio_chunks.append(pcm)
41
                    # Play immediately (16-bit mono PCM).
42
                    stream.write(np.frombuffer(pcm, dtype=np.int16))
43
                elif status == "complete":
44
                    break
45
    finally:
46
        # Always release the audio device, even if the WS errored mid-stream.
47
        stream.stop(); stream.close()
48
49
    with wave.open("streamed.wav", "wb") as wf:
50
        wf.setnchannels(1); wf.setsampwidth(2); wf.setframerate(SAMPLE_RATE)
51
        wf.writeframes(b"".join(audio_chunks))
52
    print(f"Saved streamed.wav ({len(audio_chunks)} chunks)")
53
54
asyncio.run(stream_tts("Streaming delivers audio in real-time for voice assistants and chatbots."))

1
# ci:skip — requires sounddevice + an audio device, not available in CI
2
# Plays audio chunks as they arrive AND saves the full stream to sse_output.wav.
3
# Requires: pip install requests sounddevice numpy
4
import base64
5
import json
6
import os
7
import wave
8
9
import numpy as np
10
import requests
11
import sounddevice as sd
12
13
API_KEY = os.environ["SMALLEST_API_KEY"]
14
SAMPLE_RATE = 24000
15
16
response = requests.post(
17
    "https://api.smallest.ai/waves/v1/tts/live",
18
    headers={
19
        "Authorization": f"Bearer {API_KEY}",
20
        "Content-Type": "application/json",
21
        "Accept": "text/event-stream",
22
    },
23
    json={
24
        "text": "SSE streaming is simpler to set up than WebSocket.",
25
        "voice_id": "meher",
26
        "model": "lightning_v3.1_pro",
27
        "sample_rate": SAMPLE_RATE,
28
    },
29
    stream=True,
30
)
31
response.raise_for_status()
32
33
audio_chunks = []
34
stream = sd.OutputStream(samplerate=SAMPLE_RATE, channels=1, dtype="int16")
35
try:
36
    stream.start()
37
    for line in response.iter_lines():
38
        if not line:
39
            continue
40
        decoded = line.decode()
41
        if not decoded.startswith("data: "):
42
            continue
43
        data = json.loads(decoded[6:])
44
        if data.get("done"):
45
            break
46
        if data.get("audio"):
47
            pcm = base64.b64decode(data["audio"])
48
            audio_chunks.append(pcm)
49
            # Play immediately as each chunk arrives.
50
            stream.write(np.frombuffer(pcm, dtype=np.int16))
51
finally:
52
    # Always release the audio device, even if the stream errored mid-flight.
53
    stream.stop(); stream.close()
54
55
with wave.open("sse_output.wav", "wb") as wf:
56
    wf.setnchannels(1); wf.setsampwidth(2); wf.setframerate(SAMPLE_RATE)
57
    wf.writeframes(b"".join(audio_chunks))
58
print(f"Saved sse_output.wav ({len(audio_chunks)} chunks)")