Skip to content

LiquidAI LFM3-Omni text and speech answers from audio, images, and video with the Python SDK

LiquidAI LFM3-Omni text and speech answers from audio, images, and video with the Python SDK.

Run python examples/python/lfm3_omni.py --audio question.wav --image photo.jpg. Each media flag is optional; use –prompt for a text-only request. WAV/MP3 audio uses standard input_audio parts; other formats use the audio_url extension. Video needs FFmpeg unless it is a GIF. Supply its soundtrack with –audio. Add –speech-out answer.wav for text and speech, or –audio-only for speech alone. –max-tokens counts both text tokens and speech frames (default: 1024). Increase it for longer speech.

"""
LiquidAI LFM3-Omni text and speech answers from audio, images, and video with the Python SDK.
Run `python examples/python/lfm3_omni.py --audio question.wav --image photo.jpg`.
Each media flag is optional; use --prompt for a text-only request.
WAV/MP3 audio uses standard input_audio parts; other formats use the audio_url extension.
Video needs FFmpeg unless it is a GIF. Supply its soundtrack with --audio.
Add --speech-out answer.wav for text and speech, or --audio-only for speech alone.
--max-tokens counts both text tokens and speech frames (default: 1024).
Increase it for longer speech.
"""
import argparse
import base64
import sys
from pathlib import Path
from mistralrs import ChatCompletionRequest, Runner, Which
DEFAULT_MAX_TOKENS = 1024
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--max-tokens",
type=int,
default=DEFAULT_MAX_TOKENS,
help="Positive budget for text tokens and speech frames; increase for longer speech",
)
parser.add_argument("--audio", type=Path, help="Path to an audio clip")
parser.add_argument("--image", type=Path, help="Path to an image")
parser.add_argument("--video", type=Path, help="Path to a video or animated GIF")
parser.add_argument("--speech-out", type=Path, help="Save the generated speech as WAV")
parser.add_argument(
"--audio-only", action="store_true", help="Generate speech without text"
)
parser.add_argument(
"--prompt", default="Describe the supplied media and answer any spoken question."
)
args = parser.parse_args()
if args.max_tokens <= 0:
parser.error("--max-tokens must be positive")
if args.audio_only and args.speech_out is None:
parser.error("--audio-only requires --speech-out")
output = {}
if args.speech_out is not None:
output = {
"modalities": ["audio"] if args.audio_only else ["text", "audio"],
"audio": {"format": "wav", "voice": "default"},
}
content = []
for kind, path in [("image", args.image), ("video", args.video), ("audio", args.audio)]:
if path is not None:
key = f"{kind}_url"
audio_format = path.suffix.lower().removeprefix(".")
if kind == "audio" and audio_format in {"wav", "mp3"}:
content.append(
{
"type": "input_audio",
"input_audio": {
"data": base64.b64encode(path.read_bytes()).decode("ascii"),
"format": audio_format,
},
}
)
else:
content.append({"type": key, key: {"url": path.resolve().as_uri()}})
content.append({"type": "text", "text": args.prompt})
runner = Runner(
which=Which.MultimodalPlain(model_id="LiquidAI/LFM3-Omni-2B"),
in_situ_quant="q8_0",
)
response = runner.send_chat_completion_request(
ChatCompletionRequest(
model="default",
messages=[{"role": "user", "content": content}],
max_tokens=args.max_tokens,
temperature=0.1,
**output,
)
)
choice = response.choices[0]
if choice.finish_reason == "length":
print(
"Generation reached --max-tokens; output may be incomplete. Increase --max-tokens for longer output.",
file=sys.stderr,
)
message = choice.message
if message.content:
print(message.content)
if args.speech_out is not None and message.audio is not None:
args.speech_out.write_bytes(base64.b64decode(message.audio.data))
print(f"Saved speech to {args.speech_out}")
print(response.usage)

Source: examples/python/lfm3_omni.py