[TTS][Model] support MOSS-TTS series (#3420)
Signed-off-by: Zhang <jianmusings@gmail.com> Signed-off-by: Zhang Jian <jianmusings@gmail.com> Signed-off-by: jian <jianmusings@gmail.com> Signed-off-by: Canlin Guo <961750412@qq.com> Co-authored-by: Canlin Guo <961750412@qq.com>
This commit is contained in:
1 parent
da9c45e1de
commit
b550709bba
32 files changed
+4775
-37
No files matched your search
@@ -0,0 +1,10 @@
|
||||
se001|ocean waves crashing on a rocky beach|10.0
|
||||
se002|heavy rain on a metal roof at night|8.0
|
||||
se003|a dog barking in the distance|6.0
|
||||
se004|a crackling campfire in a quiet forest|10.0
|
||||
se005|footsteps on dry leaves|6.0
|
||||
se006|thunder rolling across a wide valley|8.0
|
||||
se007|a busy coffee shop with chatter and espresso machines|10.0
|
||||
se008|wind howling through tall pine trees|8.0
|
||||
se009|a small stream flowing over pebbles|10.0
|
||||
se010|a creaking wooden door slowly opening|6.0
|
||||
@@ -0,0 +1,10 @@
|
||||
ttsd001|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Hello and welcome to the show today. [S2] Thanks for having me, it's a pleasure to be here. [S1] Let's dive right in.
|
||||
ttsd002|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Did you finish the report? [S2] Almost — I'm just polishing the conclusion. [S1] Great, send it over when you're done.
|
||||
ttsd003|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Have you tried that new restaurant downtown? [S2] Yes, the pasta is amazing. [S1] I'll book a table for Friday.
|
||||
ttsd004|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] What did you think of yesterday's meeting? [S2] Productive, but a bit long. [S1] Agreed, we should set an hour cap next time.
|
||||
ttsd005|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] The weather is beautiful today. [S2] Perfect for a hike. [S1] Want to drive out to the trails this afternoon?
|
||||
ttsd006|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] How did the demo go? [S2] Better than expected — the client signed off on the prototype. [S1] Fantastic news!
|
||||
ttsd007|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Coffee or tea this morning? [S2] Coffee please, with a splash of milk. [S1] Coming right up.
|
||||
ttsd008|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Did you read the email about the deadline change? [S2] I did, we now have until next Friday. [S1] That should give us enough buffer.
|
||||
ttsd009|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Are you joining the team dinner tonight? [S2] I'll be there around seven. [S1] See you then.
|
||||
ttsd010|prompt-wavs/spk1.wav|prompt-wavs/spk2.wav|[S1] Have we backed up the database? [S2] Last backup ran two hours ago. [S1] Good, I'll trigger a fresh snapshot before the migration.
|
||||
@@ -53,6 +53,8 @@ _TASK_TO_DATASET: dict[str, str] = {
|
||||
"voice_clone": "seed-tts",
|
||||
"default_voice": "seed-tts-text",
|
||||
"voice_design": "seed-tts-design",
|
||||
"dialogue": "ttsd",
|
||||
"sound_effect": "sound-effect",
|
||||
}
|
||||
|
||||
# Default design dataset path (bundled with the repo)
|
||||
@@ -81,6 +83,7 @@ def build_bench_args(
|
||||
output_dir: str | None,
|
||||
result_filename: str | None,
|
||||
extra_cli_args: list[str],
|
||||
output_len: int | None = None,
|
||||
) -> list[str]:
|
||||
"""Build the ``vllm bench serve --omni`` command for one (task, concurrency) run."""
|
||||
dataset_name = _TASK_TO_DATASET[task]
|
||||
@@ -139,6 +142,9 @@ def build_bench_args(
|
||||
if wer_eval:
|
||||
cmd.append("--seed-tts-wer-eval")
|
||||
|
||||
if output_len is not None:
|
||||
cmd += ["--hf-output-len", str(output_len)]
|
||||
|
||||
if output_dir or result_filename:
|
||||
out_dir = output_dir or "."
|
||||
os.makedirs(out_dir, exist_ok=True)
|
||||
@@ -238,6 +244,15 @@ def main() -> None:
|
||||
parser.add_argument("--host", default="localhost")
|
||||
parser.add_argument("--port", type=int, default=8000)
|
||||
parser.add_argument("--model-configs", default=str(_DEFAULT_MODEL_CONFIGS), help="Path to model_configs.yaml")
|
||||
parser.add_argument(
|
||||
"--output-len",
|
||||
type=int,
|
||||
default=None,
|
||||
help="Override per-request max_new_tokens forwarded to the server "
|
||||
"(seed-tts datasets default to 2048; the MOSS-TTS full talker takes "
|
||||
"this as 'max_new_frames' so smaller values cap audio length per "
|
||||
"request and keep bench runtime tractable).",
|
||||
)
|
||||
parser.add_argument("extra", nargs=argparse.REMAINDER, help="Extra args passed directly to vllm bench serve")
|
||||
args = parser.parse_args()
|
||||
|
||||
@@ -292,6 +307,7 @@ def main() -> None:
|
||||
output_dir=args.output_dir,
|
||||
result_filename=result_filename,
|
||||
extra_cli_args=args.extra or [],
|
||||
output_len=args.output_len,
|
||||
)
|
||||
result = run_one_benchmark(cmd)
|
||||
if result is not None:
|
||||
|
||||
@@ -43,3 +43,61 @@ models:
|
||||
voice_clone: {}
|
||||
default_voice:
|
||||
voice: default
|
||||
|
||||
# MOSS-TTS family. After the serving_speech.py variant-detection refactor
|
||||
# all five variants now go through OpenAI /v1/audio/speech with the
|
||||
# variant-specific request fields (ambient_sound, ref_audio_2, instructions).
|
||||
OpenMOSS-Team/MOSS-TTS:
|
||||
supported_tasks: [voice_clone]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
voice_clone: {}
|
||||
|
||||
# MOSS-TTS-v1.5: continued-training upgrade of MOSS-TTS 1.0 — same
|
||||
# MossTTSDelay-8B architecture and API, so identical wiring. Adds 31-language
|
||||
# synthesis via the `language` field and inline `[pause Xs]` markers.
|
||||
OpenMOSS-Team/MOSS-TTS-v1.5:
|
||||
supported_tasks: [voice_clone]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
voice_clone: {}
|
||||
|
||||
OpenMOSS-Team/MOSS-TTS-Realtime:
|
||||
supported_tasks: [voice_clone]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
voice_clone: {}
|
||||
|
||||
# MOSS-TTSD synthesizes [S1]/[S2] dialogue. The seed-tts voice_clone dataset
|
||||
# has single-speaker rows; bench piggybacks on it (both speakers share the
|
||||
# same ref_audio). For real dialogue evaluation use the ttsd dataset.
|
||||
OpenMOSS-Team/MOSS-TTSD-v1.0:
|
||||
supported_tasks: [voice_clone, dialogue]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
voice_clone: {}
|
||||
dialogue: {}
|
||||
|
||||
# MOSS-SoundEffect ignores text/ref_audio; expects ambient_sound +
|
||||
# duration_seconds. The sound_effect dataset module emits those.
|
||||
OpenMOSS-Team/MOSS-SoundEffect:
|
||||
supported_tasks: [sound_effect]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
sound_effect:
|
||||
duration_seconds: 10
|
||||
|
||||
# MOSS-VoiceGenerator: instructions-driven (no ref_audio). seed-tts-design
|
||||
# dataset already emits {instructions, target_text}, so voice_design works
|
||||
# out of the box once the serving validator accepts instructions.
|
||||
OpenMOSS-Team/MOSS-VoiceGenerator:
|
||||
supported_tasks: [voice_design]
|
||||
backend: openai-audio-speech
|
||||
endpoint: /v1/audio/speech
|
||||
task_extra_body:
|
||||
voice_design: {}
|
||||
Reference in new issue
Block a user