* server: add oai-compat /v1/audio/speech endpoint * add e2e pytest * server: share the s16 rounding, type speech form fields The WAV writer of the audio helper truncates while pcm_f32_to_s16 rounds, so wav and pcm differ by one LSB. Both now round. pcm_f32_to_s16 takes a pointer and a count and converts the TTS bytes without an intermediate copy. /tts and /v1/audio/speech parse their body with parse_tts_body, so multipart fields such as speed get their type before the speech conversion checks them. --------- Co-authored-by: Pascal <admin@serveurperso.com>
138 lines
4.3 KiB
Python
138 lines
4.3 KiB
Python
import pytest
|
|
import requests
|
|
import struct
|
|
from utils import *
|
|
|
|
server: ServerProcess
|
|
|
|
BODY = {
|
|
"model": "tts",
|
|
"input": "Hello there.",
|
|
"voice": "female",
|
|
"seed": 42,
|
|
"n_predict": 8,
|
|
}
|
|
|
|
|
|
@pytest.fixture(autouse=True)
|
|
def create_server():
|
|
global server
|
|
server = ServerPreset.tinylfm3omni()
|
|
|
|
|
|
def post_audio(path: str, **extra) -> requests.Response:
|
|
return requests.post(server.make_url(path), json={**BODY, **extra}, timeout=DEFAULT_REQUEST_TIMEOUT)
|
|
|
|
|
|
def assert_wav(res: requests.Response, data_size: int | None = None):
|
|
assert res.status_code == 200
|
|
assert res.headers["Content-Type"] == "audio/wav"
|
|
assert res.content[:4] == b"RIFF" and res.content[8:12] == b"WAVE"
|
|
assert struct.unpack("<H", res.content[34:36])[0] == 16 # bits per sample
|
|
pos = res.content.index(b"data")
|
|
size = struct.unpack("<I", res.content[pos + 4:pos + 8])[0]
|
|
assert size == (len(res.content) - pos - 8 if data_size is None else data_size)
|
|
assert len(res.content) > pos + 8
|
|
|
|
|
|
@pytest.mark.parametrize("path", ["/v1/audio/speech", "/audio/speech"])
|
|
def test_speech_default_format_is_wav(path: str):
|
|
server.start()
|
|
assert_wav(post_audio(path))
|
|
|
|
|
|
def test_speech_wav_same_as_tts():
|
|
server.start()
|
|
res = post_audio("/v1/audio/speech", response_format="wav")
|
|
assert_wav(res)
|
|
assert res.content == post_audio("/tts").content
|
|
|
|
|
|
def test_speech_pcm_is_signed_16_bit():
|
|
server.start()
|
|
res = post_audio("/v1/audio/speech", response_format="pcm")
|
|
assert res.status_code == 200
|
|
assert res.headers["Content-Type"] == "audio/pcm;rate=24000;encoding=signed-int;bits=16"
|
|
|
|
# /tts keeps the float samples
|
|
ref = post_audio("/tts", response_format="pcm")
|
|
assert ref.headers["Content-Type"] == "audio/pcm;rate=24000;encoding=float;bits=32"
|
|
|
|
samples = struct.unpack(f"<{len(res.content) // 2}h", res.content)
|
|
floats = struct.unpack(f"<{len(ref.content) // 4}f", ref.content)
|
|
assert len(samples) == len(floats) > 0
|
|
assert any(v != 0 for v in samples)
|
|
for s, f in zip(samples, floats):
|
|
assert abs(s - max(-1.0, min(1.0, f)) * 32767) <= 1
|
|
|
|
|
|
def test_speech_stream_pcm():
|
|
server.start()
|
|
res = post_audio("/v1/audio/speech", response_format="pcm", stream_format="audio")
|
|
assert res.status_code == 200
|
|
assert res.headers.get("Transfer-Encoding") == "chunked"
|
|
assert res.content == post_audio("/v1/audio/speech", response_format="pcm").content
|
|
|
|
|
|
def test_speech_stream_wav_has_unknown_size():
|
|
server.start()
|
|
res = post_audio("/v1/audio/speech", stream_format="audio")
|
|
assert_wav(res, data_size=0xFFFFFFFF)
|
|
|
|
|
|
def test_speech_voice():
|
|
server.start()
|
|
female = post_audio("/v1/audio/speech", voice="female")
|
|
male = post_audio("/v1/audio/speech", voice="male")
|
|
assert_wav(female)
|
|
assert_wav(male)
|
|
assert female.content != male.content
|
|
assert post_audio("/v1/audio/speech", voice={"id": "male"}).content == male.content
|
|
|
|
|
|
def test_speech_multipart():
|
|
server.start()
|
|
fields = {k: (None, str(v)) for k, v in BODY.items()}
|
|
res = requests.post(server.make_url("/v1/audio/speech"), files=fields, timeout=DEFAULT_REQUEST_TIMEOUT)
|
|
assert_wav(res)
|
|
assert res.content == post_audio("/v1/audio/speech").content
|
|
|
|
|
|
@pytest.mark.parametrize("extra", [
|
|
{"speed": 1.0},
|
|
{"instructions": "Speak slowly."},
|
|
])
|
|
def test_speech_accepted_params(extra: dict):
|
|
server.start()
|
|
assert_wav(post_audio("/v1/audio/speech", **extra))
|
|
|
|
|
|
@pytest.mark.parametrize("extra,message", [
|
|
({"response_format": "mp3"}, "response_format"),
|
|
({"response_format": "opus"}, "response_format"),
|
|
({"stream_format": "sse"}, "stream_format"),
|
|
({"speed": 1.5}, "speed"),
|
|
({"voice": "alloy"}, "unknown voice"),
|
|
({"input": ""}, "input"),
|
|
])
|
|
def test_speech_invalid_request(extra: dict, message: str):
|
|
server.start()
|
|
res = post_audio("/v1/audio/speech", **extra)
|
|
assert res.status_code == 400
|
|
assert message in res.json()["error"]["message"]
|
|
|
|
|
|
def test_speech_no_tts_model():
|
|
global server
|
|
server = ServerPreset.tinyllama2()
|
|
server.start()
|
|
assert post_audio("/v1/audio/speech").status_code == 501 # ERROR_TYPE_NOT_SUPPORTED
|
|
|
|
|
|
def test_speech_router():
|
|
global server
|
|
model = server.model_hf_repo
|
|
server = ServerPreset.router()
|
|
server.start()
|
|
assert_wav(post_audio("/v1/audio/speech", model=model))
|