Skip to content

Use the OpenAI SDK speech, transcription, and translation endpoints with LFM3-Audio

Use the OpenAI SDK speech, transcription, and translation endpoints with LFM3-Audio.

Start the server: mistralrs serve -p 1234 -m LiquidAI/LFM3-Audio-2B –isq q8_0

Generate speech: python examples/server/lfm3_audio_endpoints.py –task tts –text “Hello.” –output hello.mp3 python examples/server/lfm3_audio_endpoints.py –task tts –text “Hello.” –stream –format pcm

Transcribe or translate an uploaded clip: python examples/server/lfm3_audio_endpoints.py –task asr –audio speech.wav python examples/server/lfm3_audio_endpoints.py –task translate –audio speech.wav

Set –base-url to use another server. LFM3-Omni supports these endpoints too. MP3 is the default output format. Compressed output and speed adjustment require FFmpeg. The checkpoint has one default voice; OpenAI’s named voices are unavailable. Add –stream to ASR for transcript text events or to TTS for speech audio events. TTS streams incrementally with –format pcm at speed 1.0; other settings buffer the audio. PCM output is mono, 24 kHz, signed 16-bit little-endian audio without a file header.

"""
Use the OpenAI SDK speech, transcription, and translation endpoints with LFM3-Audio.
Start the server:
mistralrs serve -p 1234 -m LiquidAI/LFM3-Audio-2B --isq q8_0
Generate speech:
python examples/server/lfm3_audio_endpoints.py --task tts --text "Hello." --output hello.mp3
python examples/server/lfm3_audio_endpoints.py --task tts --text "Hello." --stream --format pcm
Transcribe or translate an uploaded clip:
python examples/server/lfm3_audio_endpoints.py --task asr --audio speech.wav
python examples/server/lfm3_audio_endpoints.py --task translate --audio speech.wav
Set --base-url to use another server. LFM3-Omni supports these endpoints too.
MP3 is the default output format. Compressed output and speed adjustment require FFmpeg.
The checkpoint has one default voice; OpenAI's named voices are unavailable.
Add --stream to ASR for transcript text events or to TTS for speech audio events.
TTS streams incrementally with --format pcm at speed 1.0; other settings buffer the audio.
PCM output is mono, 24 kHz, signed 16-bit little-endian audio without a file header.
"""
import argparse
import base64
import json
from pathlib import Path
from openai import OpenAI
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--base-url", default="http://localhost:1234/v1/")
parser.add_argument("--model", default="default")
parser.add_argument("--task", choices=["tts", "asr", "translate"], required=True)
parser.add_argument(
"--audio", type=Path, help="Input clip for transcription or translation"
)
parser.add_argument("--text", default="Hello.", help="Text to synthesize")
parser.add_argument("--output", type=Path, help="Output speech file")
parser.add_argument(
"--format", choices=["mp3", "opus", "aac", "flac", "wav", "pcm"], default="mp3"
)
parser.add_argument(
"--speed", type=float, default=1.0, help="Speech speed from 0.25 to 4.0"
)
parser.add_argument("--instructions", help="Additional speech-generation instructions")
parser.add_argument("--language", help="ISO 639-1 transcription language, such as en")
parser.add_argument("--prompt", help="Spelling or vocabulary context for transcription")
parser.add_argument(
"--stream", action="store_true", help="Stream ASR text or TTS audio events"
)
args = parser.parse_args()
if args.task != "tts" and args.audio is None:
parser.error("--audio is required for asr and translate")
if not 0.25 <= args.speed <= 4.0:
parser.error("--speed must be between 0.25 and 4.0")
if args.stream and args.task == "translate":
parser.error("--stream is supported for --task asr or tts")
client = OpenAI(api_key="foobar", base_url=args.base_url)
if args.task == "tts":
options = {}
if args.instructions is not None:
options["instructions"] = args.instructions
options.update(
model=args.model,
input=args.text,
voice="default",
response_format=args.format,
speed=args.speed,
)
output = args.output or Path(f"speech.{args.format}")
if args.stream:
with (
client.audio.speech.with_streaming_response.create(
stream_format="sse", **options
) as response,
output.open("wb") as audio,
):
for line in response.iter_lines():
if not line.startswith("data:"):
continue
event = json.loads(line.removeprefix("data:").strip())
if "error" in event:
raise RuntimeError(event["error"])
if event["type"] == "speech.audio.delta":
audio.write(base64.b64decode(event["audio"]))
audio.flush()
elif event["type"] == "speech.audio.done":
print(f"Usage: {event['usage']}")
else:
output.write_bytes(client.audio.speech.create(**options).content)
print(f"Saved speech to {output}")
else:
options = {}
if args.prompt is not None:
options["prompt"] = args.prompt
if args.task == "asr" and args.language is not None:
options["language"] = args.language
if args.stream:
options["stream"] = True
create = (
client.audio.transcriptions.create
if args.task == "asr"
else client.audio.translations.create
)
with args.audio.open("rb") as audio:
response = create(
model=args.model, file=audio, response_format="json", **options
)
if args.stream:
for event in response:
if event.type == "transcript.text.delta":
print(event.delta, end="", flush=True)
print()
else:
print(response.text)

Source: examples/server/lfm3_audio_endpoints.py