Datasets:

anilkeshwani
/

speech-integration-eval-sets

"""Discrete speech unit (DSU) tokenizers matching the speech-integration MLS training data.
Each tokenizer exposes `sample_rate` (expected input rate) and `encode(wav)` where `wav` is a
1-D float32 torch tensor in [-1, 1] at 16 kHz (the rate of all MLS source audio). Resampling to the
tokenizer's native rate happens inside `encode`, so callers always pass 16 kHz audio.
Tokenizer names match the directory names used under eval_data/:
    hubert_large_ll60k-layer_22, speechtokenizer-rvq_0, mimi-srvq_0, focalcodec
"""

from __future__ import annotations

import copy
import json
import os

import numpy as np
import torch
import torch.nn.functional as F
import torchaudio.functional as AF


INPUT_SR = 16_000

HUBERT_REPO = "facebook/hubert-large-ll60k"
HUBERT_LAYER = 22
KMEANS_REPO = "utter-project/SpireKMeans"
SPEECHTOKENIZER_REPO = "fnlp/SpeechTokenizer"
MIMI_REPO = "kyutai/mimi"
FOCALCODEC_REPO = "lucadellalib/focalcodec_50hz"
FOCALCODEC_REVISION = "d6d6d9524e52155c85193c2c3b8da1cf8842f019"  # weights unchanged since 2025-02-11


def _hf_file(repo_id: str, filename: str, revision: str | None = None) -> str:
    from huggingface_hub import hf_hub_download

    return hf_hub_download(repo_id, filename, revision=revision)


KMEANS_REVISION = "bb0d1a04213141f3c61912989abec9641943af60"
KMEANS_CENTERS_NPY = os.path.join(
    os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
    "downloads",
    "spire_kmeans",
    f"cluster_centers_{KMEANS_REVISION[:7]}.npy",
)


def load_kmeans_centers() -> np.ndarray:
    """Return the (5000, 1024) float32 centroids of utter-project/SpireKMeans.
    The upstream file is a joblib pickle of a sklearn MiniBatchKMeans (sklearn 1.3.2). It was
    unpickled once (it is the published k-means for this setup) and its centroids exported to .npy
    by export_spire_kmeans_centers.py, so normal runs never unpickle downloaded data.
    """
    return np.load(KMEANS_CENTERS_NPY, allow_pickle=False).astype(np.float32)


class HubertKMeansTokenizer:
    """HuBERT Large (LL-60k) layer-22 features + 5000-cluster k-means (SpireKMeans).
    Mirrors sardalign.utils.features.SimpleHubertFeaturizer + sardalign.dump_km_label.ApplyKmeans:
    the whole waveform is layer-normalized (task.cfg.normalize=True for HuBERT Large), then fed
    unpadded; layer 22 output (before the final encoder LayerNorm) is assigned to the nearest centroid.
    """

    name = "hubert_large_ll60k-layer_22"
    sample_rate = 16_000

    def __init__(self, device: torch.device):
        from transformers import HubertModel

        self.device = device
        model = HubertModel.from_pretrained(HUBERT_REPO, torch_dtype=torch.float32)
        # Drop unused top layers. Keep one extra: the encoder applies its final LayerNorm to the last
        # hidden state, so with exactly 22 layers hidden_states[22] would be normalized.
        model.encoder.layers = model.encoder.layers[: HUBERT_LAYER + 1]
        self.model = model.eval().to(device)
        C = torch.from_numpy(load_kmeans_centers().T).to(device)  # (D, K)
        self.C = C
        self.Cnorm = (C**2).sum(0, keepdim=True)

    @torch.inference_mode()
    def encode(self, wav: torch.Tensor) -> list[int]:
        x = wav.to(self.device, torch.float32)
        x = F.layer_norm(x, x.shape).view(1, -1)
        out = self.model(x, output_hidden_states=True)
        feat = out.hidden_states[HUBERT_LAYER].squeeze(0)  # (T, 1024)
        dist = feat.pow(2).sum(1, keepdim=True) - 2 * feat @ self.C + self.Cnorm
        return dist.argmin(dim=1).cpu().tolist()


class SpeechTokenizerRVQ0:
    """SpeechTokenizer (speechtokenizer_hubert_avg), first RVQ codebook (1024 entries), 50 Hz."""

    name = "speechtokenizer-rvq_0"
    sample_rate = 16_000

    def __init__(self, device: torch.device):
        from speechtokenizer import SpeechTokenizer

        cfg = _hf_file(SPEECHTOKENIZER_REPO, "speechtokenizer_hubert_avg/config.json")
        ckpt = _hf_file(SPEECHTOKENIZER_REPO, "speechtokenizer_hubert_avg/SpeechTokenizer.pt")
        self.model = SpeechTokenizer.load_from_checkpoint(cfg, ckpt).eval().to(device)
        assert self.model.sample_rate == self.sample_rate
        self.device = device

    @torch.inference_mode()
    def encode(self, wav: torch.Tensor) -> list[int]:
        x = wav.to(self.device, torch.float32).view(1, 1, -1)
        codes = self.model.encode(x, n_q=1)  # (n_q, B, T)
        return codes[0, 0].cpu().tolist()


class MimiSRVQ0:
    """Kyutai Mimi (transformers MimiModel), split-RVQ level 0 (semantic codebook, 2048), 12.5 Hz."""

    name = "mimi-srvq_0"
    sample_rate = 24_000

    def __init__(self, device: torch.device, resample: str = "torchaudio"):
        from transformers import MimiModel

        self.model = MimiModel.from_pretrained(MIMI_REPO).eval().to(device)
        self.device = device
        self.resample = resample

    def _resample(self, wav: torch.Tensor) -> torch.Tensor:
        if self.resample == "torchaudio":
            return AF.resample(wav, INPUT_SR, self.sample_rate)
        if self.resample.startswith("librosa"):
            import librosa

            res_type = self.resample.split(":", 1)[1] if ":" in self.resample else "soxr_hq"
            y = librosa.resample(wav.numpy(), orig_sr=INPUT_SR, target_sr=self.sample_rate, res_type=res_type)
            return torch.from_numpy(y)
        raise ValueError(self.resample)

    @torch.inference_mode()
    def encode(self, wav: torch.Tensor) -> list[int]:
        x = self._resample(wav.float().cpu()).to(self.device).view(1, 1, -1)
        codes = self.model.encode(x, num_quantizers=8).audio_codes  # (B, n_q, T)
        return codes[0, 0].cpu().tolist()


class FocalCodec50Hz:
    """FocalCodec 50 Hz (single binary spherical quantizer codebook, 8192 entries)."""

    name = "focalcodec"
    sample_rate = 16_000

    def __init__(self, device: torch.device, cpu_above_s: float = 40.0):
        from focalcodec import FocalCodec
        from safetensors.torch import load_file

        # Same as FocalCodec.from_pretrained(FOCALCODEC_REPO) in focalcodec 0.0.1, but cache-aware
        # (from_config calls list_repo_files, which fails with HF_HUB_OFFLINE=1)
        with open(_hf_file(FOCALCODEC_REPO, "config.json", FOCALCODEC_REVISION)) as f:
            model = FocalCodec(**json.load(f))
        model.load_state_dict(load_file(_hf_file(FOCALCODEC_REPO, "model.safetensors", FOCALCODEC_REVISION)))
        self.model = model.eval().requires_grad_(False).to(device)
        # focalcodec 0.0.1 (pinned commit 68be6ce, the code current when the MLS tokens were made)
        # exposes `sample_rate`; later versions renamed it to `sample_rate_input`
        sr = getattr(self.model, "sample_rate_input", None) or self.model.sample_rate
        assert sr == self.sample_rate, sr
        self.device = device
        self.cpu_above_s = cpu_above_s
        self._cpu_model = None

    @torch.inference_mode()
    def encode(self, wav: torch.Tensor) -> list[int]:
        # WavLM attention needs ~5 GiB extra GPU memory at 100 s, so long inputs go to a CPU copy of
        # the model (CPU and GPU fp32 gave identical tokens in testing)
        if self.device.type == "cuda" and wav.numel() > self.cpu_above_s * self.sample_rate:
            if self._cpu_model is None:
                self._cpu_model = copy.deepcopy(self.model).cpu()
            model, device = self._cpu_model, torch.device("cpu")
        else:
            model, device = self.model, self.device
        x = wav.to(device, torch.float32).view(1, -1)
        toks = model.sig_to_toks(x)  # (B, T)
        return toks[0].cpu().tolist()


TOKENIZERS = {
    cls.name: cls for cls in (HubertKMeansTokenizer, SpeechTokenizerRVQ0, MimiSRVQ0, FocalCodec50Hz)
}


def get_device() -> torch.device:
    if torch.cuda.is_available():
        visible = os.environ.get("CUDA_VISIBLE_DEVICES")
        if visible != "3":
            raise RuntimeError(f"Refusing to run on GPUs other than 3; CUDA_VISIBLE_DEVICES={visible!r}")
        return torch.device("cuda:0")
    return torch.device("cpu")

	"""Discrete speech unit (DSU) tokenizers matching the speech-integration MLS training data.

	Each tokenizer exposes `sample_rate` (expected input rate) and `encode(wav)` where `wav` is a
	1-D float32 torch tensor in [-1, 1] at 16 kHz (the rate of all MLS source audio). Resampling to the
	tokenizer's native rate happens inside `encode`, so callers always pass 16 kHz audio.

	Tokenizer names match the directory names used under eval_data/:
	hubert_large_ll60k-layer_22, speechtokenizer-rvq_0, mimi-srvq_0, focalcodec
	"""

	from __future__ import annotations

	import copy
	import json
	import os

	import numpy as np
	import torch
	import torch.nn.functional as F
	import torchaudio.functional as AF


	INPUT_SR = 16_000

	HUBERT_REPO = "facebook/hubert-large-ll60k"
	HUBERT_LAYER = 22
	KMEANS_REPO = "utter-project/SpireKMeans"
	SPEECHTOKENIZER_REPO = "fnlp/SpeechTokenizer"
	MIMI_REPO = "kyutai/mimi"
	FOCALCODEC_REPO = "lucadellalib/focalcodec_50hz"
	FOCALCODEC_REVISION = "d6d6d9524e52155c85193c2c3b8da1cf8842f019" # weights unchanged since 2025-02-11


	def _hf_file(repo_id: str, filename: str, revision: str \| None = None) -> str:
	from huggingface_hub import hf_hub_download

	return hf_hub_download(repo_id, filename, revision=revision)


	KMEANS_REVISION = "bb0d1a04213141f3c61912989abec9641943af60"
	KMEANS_CENTERS_NPY = os.path.join(
	os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
	"downloads",
	"spire_kmeans",
	f"cluster_centers_{KMEANS_REVISION[:7]}.npy",
	)


	def load_kmeans_centers() -> np.ndarray:
	"""Return the (5000, 1024) float32 centroids of utter-project/SpireKMeans.

	The upstream file is a joblib pickle of a sklearn MiniBatchKMeans (sklearn 1.3.2). It was
	unpickled once (it is the published k-means for this setup) and its centroids exported to .npy
	by export_spire_kmeans_centers.py, so normal runs never unpickle downloaded data.
	"""
	return np.load(KMEANS_CENTERS_NPY, allow_pickle=False).astype(np.float32)


	class HubertKMeansTokenizer:
	"""HuBERT Large (LL-60k) layer-22 features + 5000-cluster k-means (SpireKMeans).

	Mirrors sardalign.utils.features.SimpleHubertFeaturizer + sardalign.dump_km_label.ApplyKmeans:
	the whole waveform is layer-normalized (task.cfg.normalize=True for HuBERT Large), then fed
	unpadded; layer 22 output (before the final encoder LayerNorm) is assigned to the nearest centroid.
	"""

	name = "hubert_large_ll60k-layer_22"
	sample_rate = 16_000

	def __init__(self, device: torch.device):
	from transformers import HubertModel

	self.device = device
	model = HubertModel.from_pretrained(HUBERT_REPO, torch_dtype=torch.float32)
	# Drop unused top layers. Keep one extra: the encoder applies its final LayerNorm to the last
	# hidden state, so with exactly 22 layers hidden_states[22] would be normalized.
	model.encoder.layers = model.encoder.layers[: HUBERT_LAYER + 1]
	self.model = model.eval().to(device)
	C = torch.from_numpy(load_kmeans_centers().T).to(device) # (D, K)
	self.C = C
	self.Cnorm = (C**2).sum(0, keepdim=True)

	@torch.inference_mode()
	def encode(self, wav: torch.Tensor) -> list[int]:
	x = wav.to(self.device, torch.float32)
	x = F.layer_norm(x, x.shape).view(1, -1)
	out = self.model(x, output_hidden_states=True)
	feat = out.hidden_states[HUBERT_LAYER].squeeze(0) # (T, 1024)
	dist = feat.pow(2).sum(1, keepdim=True) - 2 * feat @ self.C + self.Cnorm
	return dist.argmin(dim=1).cpu().tolist()


	class SpeechTokenizerRVQ0:
	"""SpeechTokenizer (speechtokenizer_hubert_avg), first RVQ codebook (1024 entries), 50 Hz."""

	name = "speechtokenizer-rvq_0"
	sample_rate = 16_000

	def __init__(self, device: torch.device):
	from speechtokenizer import SpeechTokenizer

	cfg = _hf_file(SPEECHTOKENIZER_REPO, "speechtokenizer_hubert_avg/config.json")
	ckpt = _hf_file(SPEECHTOKENIZER_REPO, "speechtokenizer_hubert_avg/SpeechTokenizer.pt")
	self.model = SpeechTokenizer.load_from_checkpoint(cfg, ckpt).eval().to(device)
	assert self.model.sample_rate == self.sample_rate
	self.device = device

	@torch.inference_mode()
	def encode(self, wav: torch.Tensor) -> list[int]:
	x = wav.to(self.device, torch.float32).view(1, 1, -1)
	codes = self.model.encode(x, n_q=1) # (n_q, B, T)
	return codes[0, 0].cpu().tolist()


	class MimiSRVQ0:
	"""Kyutai Mimi (transformers MimiModel), split-RVQ level 0 (semantic codebook, 2048), 12.5 Hz."""

	name = "mimi-srvq_0"
	sample_rate = 24_000

	def __init__(self, device: torch.device, resample: str = "torchaudio"):
	from transformers import MimiModel

	self.model = MimiModel.from_pretrained(MIMI_REPO).eval().to(device)
	self.device = device
	self.resample = resample

	def _resample(self, wav: torch.Tensor) -> torch.Tensor:
	if self.resample == "torchaudio":
	return AF.resample(wav, INPUT_SR, self.sample_rate)
	if self.resample.startswith("librosa"):
	import librosa

	res_type = self.resample.split(":", 1)[1] if ":" in self.resample else "soxr_hq"
	y = librosa.resample(wav.numpy(), orig_sr=INPUT_SR, target_sr=self.sample_rate, res_type=res_type)
	return torch.from_numpy(y)
	raise ValueError(self.resample)

	@torch.inference_mode()
	def encode(self, wav: torch.Tensor) -> list[int]:
	x = self._resample(wav.float().cpu()).to(self.device).view(1, 1, -1)
	codes = self.model.encode(x, num_quantizers=8).audio_codes # (B, n_q, T)
	return codes[0, 0].cpu().tolist()


	class FocalCodec50Hz:
	"""FocalCodec 50 Hz (single binary spherical quantizer codebook, 8192 entries)."""

	name = "focalcodec"
	sample_rate = 16_000

	def __init__(self, device: torch.device, cpu_above_s: float = 40.0):
	from focalcodec import FocalCodec
	from safetensors.torch import load_file

	# Same as FocalCodec.from_pretrained(FOCALCODEC_REPO) in focalcodec 0.0.1, but cache-aware
	# (from_config calls list_repo_files, which fails with HF_HUB_OFFLINE=1)
	with open(_hf_file(FOCALCODEC_REPO, "config.json", FOCALCODEC_REVISION)) as f:
	model = FocalCodec(**json.load(f))
	model.load_state_dict(load_file(_hf_file(FOCALCODEC_REPO, "model.safetensors", FOCALCODEC_REVISION)))
	self.model = model.eval().requires_grad_(False).to(device)
	# focalcodec 0.0.1 (pinned commit 68be6ce, the code current when the MLS tokens were made)
	# exposes `sample_rate`; later versions renamed it to `sample_rate_input`
	sr = getattr(self.model, "sample_rate_input", None) or self.model.sample_rate
	assert sr == self.sample_rate, sr
	self.device = device
	self.cpu_above_s = cpu_above_s
	self._cpu_model = None

	@torch.inference_mode()
	def encode(self, wav: torch.Tensor) -> list[int]:
	# WavLM attention needs ~5 GiB extra GPU memory at 100 s, so long inputs go to a CPU copy of
	# the model (CPU and GPU fp32 gave identical tokens in testing)
	if self.device.type == "cuda" and wav.numel() > self.cpu_above_s * self.sample_rate:
	if self._cpu_model is None:
	self._cpu_model = copy.deepcopy(self.model).cpu()
	model, device = self._cpu_model, torch.device("cpu")
	else:
	model, device = self.model, self.device
	x = wav.to(device, torch.float32).view(1, -1)
	toks = model.sig_to_toks(x) # (B, T)
	return toks[0].cpu().tolist()


	TOKENIZERS = {
	cls.name: cls for cls in (HubertKMeansTokenizer, SpeechTokenizerRVQ0, MimiSRVQ0, FocalCodec50Hz)
	}


	def get_device() -> torch.device:
	if torch.cuda.is_available():
	visible = os.environ.get("CUDA_VISIBLE_DEVICES")
	if visible != "3":
	raise RuntimeError(f"Refusing to run on GPUs other than 3; CUDA_VISIBLE_DEVICES={visible!r}")
	return torch.device("cuda:0")
	return torch.device("cpu")