.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "tutorials/forced_alignment_tutorial.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note Click :ref:`here ` to download the full example code .. rst-class:: sphx-glr-example-title .. _sphx_glr_tutorials_forced_alignment_tutorial.py: Forced Alignment with Wav2Vec2 ============================== **Author**: `Moto Hira `__ This tutorial shows how to align transcript to speech with ``torchaudio``, using CTC segmentation algorithm described in `CTC-Segmentation of Large Corpora for German End-to-end Speech Recognition `__. .. GENERATED FROM PYTHON SOURCE LINES 13-25 .. code-block:: default import torch import torchaudio print(torch.__version__) print(torchaudio.__version__) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(device) .. rst-class:: sphx-glr-script-out .. code-block:: none 1.13.0 0.13.0 cpu .. GENERATED FROM PYTHON SOURCE LINES 26-39 Overview -------- The process of alignment looks like the following. 1. Estimate the frame-wise label probability from audio waveform 2. Generate the trellis matrix which represents the probability of labels aligned at time step. 3. Find the most likely path from the trellis matrix. In this example, we use ``torchaudio``\ ’s ``Wav2Vec2`` model for acoustic feature extraction. .. GENERATED FROM PYTHON SOURCE LINES 42-47 Preparation ----------- First we import the necessary packages, and fetch data that we work on. .. GENERATED FROM PYTHON SOURCE LINES 47-63 .. code-block:: default # %matplotlib inline from dataclasses import dataclass import IPython import matplotlib import matplotlib.pyplot as plt matplotlib.rcParams["figure.figsize"] = [16.0, 4.8] torch.random.manual_seed(0) SPEECH_FILE = torchaudio.utils.download_asset("tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav") .. GENERATED FROM PYTHON SOURCE LINES 64-80 Generate frame-wise label probability ------------------------------------- The first step is to generate the label class porbability of each aduio frame. We can use a Wav2Vec2 model that is trained for ASR. Here we use :py:func:`torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H`. ``torchaudio`` provides easy access to pretrained models with associated labels. .. note:: In the subsequent sections, we will compute the probability in log-domain to avoid numerical instability. For this purpose, we normalize the ``emission`` with :py:func:`torch.log_softmax`. .. GENERATED FROM PYTHON SOURCE LINES 80-91 .. code-block:: default bundle = torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H model = bundle.get_model().to(device) labels = bundle.get_labels() with torch.inference_mode(): waveform, _ = torchaudio.load(SPEECH_FILE) emissions, _ = model(waveform.to(device)) emissions = torch.log_softmax(emissions, dim=-1) emission = emissions[0].cpu().detach() .. GENERATED FROM PYTHON SOURCE LINES 92-94 Visualization ############################################################################### .. GENERATED FROM PYTHON SOURCE LINES 94-103 .. code-block:: default print(labels) plt.imshow(emission.T) plt.colorbar() plt.title("Frame-wise class probability") plt.xlabel("Time") plt.ylabel("Labels") plt.show() .. image-sg:: /tutorials/images/sphx_glr_forced_alignment_tutorial_001.png :alt: Frame-wise class probability :srcset: /tutorials/images/sphx_glr_forced_alignment_tutorial_001.png :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out .. code-block:: none ('-', '|', 'E', 'T', 'A', 'O', 'N', 'I', 'H', 'S', 'R', 'D', 'L', 'U', 'M', 'W', 'C', 'F', 'G', 'Y', 'P', 'B', 'V', 'K', "'", 'X', 'J', 'Q', 'Z') .. GENERATED FROM PYTHON SOURCE LINES 104-140 Generate alignment probability (trellis) ---------------------------------------- From the emission matrix, next we generate the trellis which represents the probability of transcript labels occur at each time frame. Trellis is 2D matrix with time axis and label axis. The label axis represents the transcript that we are aligning. In the following, we use :math:`t` to denote the index in time axis and :math:`j` to denote the index in label axis. :math:`c_j` represents the label at label index :math:`j`. To generate, the probability of time step :math:`t+1`, we look at the trellis from time step :math:`t` and emission at time step :math:`t+1`. There are two path to reach to time step :math:`t+1` with label :math:`c_{j+1}`. The first one is the case where the label was :math:`c_{j+1}` at :math:`t` and there was no label change from :math:`t` to :math:`t+1`. The other case is where the label was :math:`c_j` at :math:`t` and it transitioned to the next label :math:`c_{j+1}` at :math:`t+1`. The follwoing diagram illustrates this transition. .. image:: https://download.pytorch.org/torchaudio/tutorial-assets/ctc-forward.png Since we are looking for the most likely transitions, we take the more likely path for the value of :math:`k_{(t+1, j+1)}`, that is :math:`k_{(t+1, j+1)} = max( k_{(t, j)} p(t+1, c_{j+1}), k_{(t, j+1)} p(t+1, repeat) )` where :math:`k` represents is trellis matrix, and :math:`p(t, c_j)` represents the probability of label :math:`c_j` at time step :math:`t`. :math:`repeat` represents the blank token from CTC formulation. (For the detail of CTC algorithm, please refer to the *Sequence Modeling with CTC* [`distill.pub `__]) .. GENERATED FROM PYTHON SOURCE LINES 140-173 .. code-block:: default transcript = "I|HAD|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT" dictionary = {c: i for i, c in enumerate(labels)} tokens = [dictionary[c] for c in transcript] print(list(zip(transcript, tokens))) def get_trellis(emission, tokens, blank_id=0): num_frame = emission.size(0) num_tokens = len(tokens) # Trellis has extra diemsions for both time axis and tokens. # The extra dim for tokens represents (start-of-sentence) # The extra dim for time axis is for simplification of the code. trellis = torch.empty((num_frame + 1, num_tokens + 1)) trellis[0, 0] = 0 trellis[1:, 0] = torch.cumsum(emission[:, 0], 0) trellis[0, -num_tokens:] = -float("inf") trellis[-num_tokens:, 0] = float("inf") for t in range(num_frame): trellis[t + 1, 1:] = torch.maximum( # Score for staying at the same token trellis[t, 1:] + emission[t, blank_id], # Score for changing to the next token trellis[t, :-1] + emission[t, tokens], ) return trellis trellis = get_trellis(emission, tokens) .. rst-class:: sphx-glr-script-out .. code-block:: none [('I', 7), ('|', 1), ('H', 8), ('A', 4), ('D', 11), ('|', 1), ('T', 3), ('H', 8), ('A', 4), ('T', 3), ('|', 1), ('C', 16), ('U', 13), ('R', 10), ('I', 7), ('O', 5), ('S', 9), ('I', 7), ('T', 3), ('Y', 19), ('|', 1), ('B', 21), ('E', 2), ('S', 9), ('I', 7), ('D', 11), ('E', 2), ('|', 1), ('M', 14), ('E', 2), ('|', 1), ('A', 4), ('T', 3), ('|', 1), ('T', 3), ('H', 8), ('I', 7), ('S', 9), ('|', 1), ('M', 14), ('O', 5), ('M', 14), ('E', 2), ('N', 6), ('T', 3)] .. GENERATED FROM PYTHON SOURCE LINES 174-176 Visualization ############################################################################### .. GENERATED FROM PYTHON SOURCE LINES 176-181 .. code-block:: default plt.imshow(trellis[1:, 1:].T, origin="lower") plt.annotate("- Inf", (trellis.size(1) / 5, trellis.size(1) / 1.5)) plt.colorbar() plt.show() .. image-sg:: /tutorials/images/sphx_glr_forced_alignment_tutorial_002.png :alt: forced alignment tutorial :srcset: /tutorials/images/sphx_glr_forced_alignment_tutorial_002.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 182-185 In the above visualization, we can see that there is a trace of high probability crossing the matrix diagonally. .. GENERATED FROM PYTHON SOURCE LINES 188-207 Find the most likely path (backtracking) ---------------------------------------- Once the trellis is generated, we will traverse it following the elements with high probability. We will start from the last label index with the time step of highest probability, then, we traverse back in time, picking stay (:math:`c_j \rightarrow c_j`) or transition (:math:`c_j \rightarrow c_{j+1}`), based on the post-transition probability :math:`k_{t, j} p(t+1, c_{j+1})` or :math:`k_{t, j+1} p(t+1, repeat)`. Transition is done once the label reaches the beginning. The trellis matrix is used for path-finding, but for the final probability of each segment, we take the frame-wise probability from emission matrix. .. GENERATED FROM PYTHON SOURCE LINES 207-257 .. code-block:: default @dataclass class Point: token_index: int time_index: int score: float def backtrack(trellis, emission, tokens, blank_id=0): # Note: # j and t are indices for trellis, which has extra dimensions # for time and tokens at the beginning. # When referring to time frame index `T` in trellis, # the corresponding index in emission is `T-1`. # Similarly, when referring to token index `J` in trellis, # the corresponding index in transcript is `J-1`. j = trellis.size(1) - 1 t_start = torch.argmax(trellis[:, j]).item() path = [] for t in range(t_start, 0, -1): # 1. Figure out if the current position was stay or change # Note (again): # `emission[J-1]` is the emission at time frame `J` of trellis dimension. # Score for token staying the same from time frame J-1 to T. stayed = trellis[t - 1, j] + emission[t - 1, blank_id] # Score for token changing from C-1 at T-1 to J at T. changed = trellis[t - 1, j - 1] + emission[t - 1, tokens[j - 1]] # 2. Store the path with frame-wise probability. prob = emission[t - 1, tokens[j - 1] if changed > stayed else 0].exp().item() # Return token index and time index in non-trellis coordinate. path.append(Point(j - 1, t - 1, prob)) # 3. Update the token if changed > stayed: j -= 1 if j == 0: break else: raise ValueError("Failed to align") return path[::-1] path = backtrack(trellis, emission, tokens) for p in path: print(p) .. rst-class:: sphx-glr-script-out .. code-block:: none Point(token_index=0, time_index=30, score=0.9999842643737793) Point(token_index=0, time_index=31, score=0.9846959710121155) Point(token_index=0, time_index=32, score=0.9999706745147705) Point(token_index=0, time_index=33, score=0.15398800373077393) Point(token_index=1, time_index=34, score=0.9999173879623413) Point(token_index=1, time_index=35, score=0.6080105900764465) Point(token_index=2, time_index=36, score=0.9997721314430237) Point(token_index=2, time_index=37, score=0.9997127652168274) Point(token_index=3, time_index=38, score=0.9999358654022217) Point(token_index=3, time_index=39, score=0.9861598610877991) Point(token_index=4, time_index=40, score=0.9238577485084534) Point(token_index=4, time_index=41, score=0.9257409572601318) Point(token_index=4, time_index=42, score=0.01566212624311447) Point(token_index=5, time_index=43, score=0.9998378753662109) Point(token_index=6, time_index=44, score=0.9988443851470947) Point(token_index=6, time_index=45, score=0.10147196799516678) Point(token_index=7, time_index=46, score=0.9999427795410156) Point(token_index=7, time_index=47, score=0.9999943971633911) Point(token_index=8, time_index=48, score=0.9979604482650757) Point(token_index=8, time_index=49, score=0.03603758662939072) Point(token_index=8, time_index=50, score=0.06163697689771652) Point(token_index=9, time_index=51, score=4.3358733819331974e-05) Point(token_index=10, time_index=52, score=0.9999799728393555) Point(token_index=10, time_index=53, score=0.9967100620269775) Point(token_index=10, time_index=54, score=0.9999256134033203) Point(token_index=11, time_index=55, score=0.9999983310699463) Point(token_index=11, time_index=56, score=0.9990687966346741) Point(token_index=11, time_index=57, score=0.9999996423721313) Point(token_index=11, time_index=58, score=0.9999996423721313) Point(token_index=11, time_index=59, score=0.8457656502723694) Point(token_index=12, time_index=60, score=0.9999996423721313) Point(token_index=12, time_index=61, score=0.9996013045310974) Point(token_index=13, time_index=62, score=0.999998927116394) Point(token_index=13, time_index=63, score=0.0035240992438048124) Point(token_index=13, time_index=64, score=1.0) Point(token_index=13, time_index=65, score=1.0) Point(token_index=14, time_index=66, score=0.9999915361404419) Point(token_index=14, time_index=67, score=0.9971572160720825) Point(token_index=14, time_index=68, score=0.9999990463256836) Point(token_index=14, time_index=69, score=0.9999992847442627) Point(token_index=14, time_index=70, score=0.9999997615814209) Point(token_index=14, time_index=71, score=0.9999998807907104) Point(token_index=14, time_index=72, score=0.9999880790710449) Point(token_index=14, time_index=73, score=0.011424383148550987) Point(token_index=15, time_index=74, score=0.9999977350234985) Point(token_index=15, time_index=75, score=0.9996136426925659) Point(token_index=15, time_index=76, score=0.999998927116394) Point(token_index=15, time_index=77, score=0.9727645516395569) Point(token_index=16, time_index=78, score=0.999998927116394) Point(token_index=16, time_index=79, score=0.9949339032173157) Point(token_index=16, time_index=80, score=0.999998927116394) Point(token_index=16, time_index=81, score=0.9999121427536011) Point(token_index=17, time_index=82, score=0.9999774694442749) Point(token_index=17, time_index=83, score=0.6575831770896912) Point(token_index=17, time_index=84, score=0.9984298348426819) Point(token_index=18, time_index=85, score=0.9999876022338867) Point(token_index=18, time_index=86, score=0.9993745684623718) Point(token_index=18, time_index=87, score=0.9999988079071045) Point(token_index=18, time_index=88, score=0.10426424443721771) Point(token_index=19, time_index=89, score=0.9999969005584717) Point(token_index=19, time_index=90, score=0.39782625436782837) Point(token_index=20, time_index=91, score=0.9999932050704956) Point(token_index=20, time_index=92, score=1.6989914684018004e-06) Point(token_index=20, time_index=93, score=0.986134946346283) Point(token_index=21, time_index=94, score=0.9999960660934448) Point(token_index=21, time_index=95, score=0.9992731213569641) Point(token_index=21, time_index=96, score=0.9993410706520081) Point(token_index=22, time_index=97, score=0.9999983310699463) Point(token_index=22, time_index=98, score=0.9999971389770508) Point(token_index=22, time_index=99, score=0.9999998807907104) Point(token_index=22, time_index=100, score=0.9999995231628418) Point(token_index=23, time_index=101, score=0.9999732971191406) Point(token_index=23, time_index=102, score=0.9983221888542175) Point(token_index=23, time_index=103, score=0.9999991655349731) Point(token_index=23, time_index=104, score=0.9999996423721313) Point(token_index=23, time_index=105, score=0.9999998807907104) Point(token_index=23, time_index=106, score=1.0) Point(token_index=23, time_index=107, score=0.9998629093170166) Point(token_index=24, time_index=108, score=0.9999980926513672) Point(token_index=24, time_index=109, score=0.9988582134246826) Point(token_index=25, time_index=110, score=0.9999798536300659) Point(token_index=25, time_index=111, score=0.8573304414749146) Point(token_index=26, time_index=112, score=0.9999847412109375) Point(token_index=26, time_index=113, score=0.9870284795761108) Point(token_index=26, time_index=114, score=1.9046094166697003e-05) Point(token_index=27, time_index=115, score=0.9999794960021973) Point(token_index=27, time_index=116, score=0.9998254776000977) Point(token_index=28, time_index=117, score=0.9999990463256836) Point(token_index=28, time_index=118, score=0.9999732971191406) Point(token_index=28, time_index=119, score=0.0009001801954582334) Point(token_index=29, time_index=120, score=0.999347984790802) Point(token_index=29, time_index=121, score=0.9975457787513733) Point(token_index=29, time_index=122, score=0.00030502272420562804) Point(token_index=30, time_index=123, score=0.9999344348907471) Point(token_index=30, time_index=124, score=6.078777914808597e-06) Point(token_index=31, time_index=125, score=0.9833180904388428) Point(token_index=32, time_index=126, score=0.9974581599235535) Point(token_index=32, time_index=127, score=0.0008237161673605442) Point(token_index=33, time_index=128, score=0.996515154838562) Point(token_index=33, time_index=129, score=0.017463242635130882) Point(token_index=34, time_index=130, score=0.9989172220230103) Point(token_index=35, time_index=131, score=0.9999697208404541) Point(token_index=35, time_index=132, score=0.9999842643737793) Point(token_index=36, time_index=133, score=0.9997640252113342) Point(token_index=36, time_index=134, score=0.5097627639770508) Point(token_index=37, time_index=135, score=0.9998301267623901) Point(token_index=37, time_index=136, score=0.08528485894203186) Point(token_index=37, time_index=137, score=0.00407379399985075) Point(token_index=38, time_index=138, score=0.9999815225601196) Point(token_index=38, time_index=139, score=0.012051556259393692) Point(token_index=38, time_index=140, score=0.9999980926513672) Point(token_index=38, time_index=141, score=0.0005778882768936455) Point(token_index=39, time_index=142, score=0.9999067783355713) Point(token_index=39, time_index=143, score=0.9999960660934448) Point(token_index=39, time_index=144, score=0.9999980926513672) Point(token_index=40, time_index=145, score=0.9999916553497314) Point(token_index=40, time_index=146, score=0.9971166849136353) Point(token_index=40, time_index=147, score=0.9981802701950073) Point(token_index=41, time_index=148, score=0.9999310970306396) Point(token_index=41, time_index=149, score=0.98795086145401) Point(token_index=41, time_index=150, score=0.9997627139091492) Point(token_index=42, time_index=151, score=0.9999535083770752) Point(token_index=43, time_index=152, score=0.9999715089797974) Point(token_index=44, time_index=153, score=0.6811069846153259) .. GENERATED FROM PYTHON SOURCE LINES 258-260 Visualization ############################################################################### .. GENERATED FROM PYTHON SOURCE LINES 260-272 .. code-block:: default def plot_trellis_with_path(trellis, path): # To plot trellis with path, we take advantage of 'nan' value trellis_with_path = trellis.clone() for _, p in enumerate(path): trellis_with_path[p.time_index, p.token_index] = float("nan") plt.imshow(trellis_with_path[1:, 1:].T, origin="lower") plot_trellis_with_path(trellis, path) plt.title("The path found by backtracking") plt.show() .. image-sg:: /tutorials/images/sphx_glr_forced_alignment_tutorial_003.png :alt: The path found by backtracking :srcset: /tutorials/images/sphx_glr_forced_alignment_tutorial_003.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 273-279 Looking good. Now this path contains repetations for the same labels, so let’s merge them to make it close to the original transcript. When merging the multiple path points, we simply take the average probability for the merged segments. .. GENERATED FROM PYTHON SOURCE LINES 279-321 .. code-block:: default # Merge the labels @dataclass class Segment: label: str start: int end: int score: float def __repr__(self): return f"{self.label}\t({self.score:4.2f}): [{self.start:5d}, {self.end:5d})" @property def length(self): return self.end - self.start def merge_repeats(path): i1, i2 = 0, 0 segments = [] while i1 < len(path): while i2 < len(path) and path[i1].token_index == path[i2].token_index: i2 += 1 score = sum(path[k].score for k in range(i1, i2)) / (i2 - i1) segments.append( Segment( transcript[path[i1].token_index], path[i1].time_index, path[i2 - 1].time_index + 1, score, ) ) i1 = i2 return segments segments = merge_repeats(path) for seg in segments: print(seg) .. rst-class:: sphx-glr-script-out .. code-block:: none I (0.78): [ 30, 34) | (0.80): [ 34, 36) H (1.00): [ 36, 38) A (0.99): [ 38, 40) D (0.62): [ 40, 43) | (1.00): [ 43, 44) T (0.55): [ 44, 46) H (1.00): [ 46, 48) A (0.37): [ 48, 51) T (0.00): [ 51, 52) | (1.00): [ 52, 55) C (0.97): [ 55, 60) U (1.00): [ 60, 62) R (0.75): [ 62, 66) I (0.88): [ 66, 74) O (0.99): [ 74, 78) S (1.00): [ 78, 82) I (0.89): [ 82, 85) T (0.78): [ 85, 89) Y (0.70): [ 89, 91) | (0.66): [ 91, 94) B (1.00): [ 94, 97) E (1.00): [ 97, 101) S (1.00): [ 101, 108) I (1.00): [ 108, 110) D (0.93): [ 110, 112) E (0.66): [ 112, 115) | (1.00): [ 115, 117) M (0.67): [ 117, 120) E (0.67): [ 120, 123) | (0.50): [ 123, 125) A (0.98): [ 125, 126) T (0.50): [ 126, 128) | (0.51): [ 128, 130) T (1.00): [ 130, 131) H (1.00): [ 131, 133) I (0.75): [ 133, 135) S (0.36): [ 135, 138) | (0.50): [ 138, 142) M (1.00): [ 142, 145) O (1.00): [ 145, 148) M (1.00): [ 148, 151) E (1.00): [ 151, 152) N (1.00): [ 152, 153) T (0.68): [ 153, 154) .. GENERATED FROM PYTHON SOURCE LINES 322-324 Visualization ############################################################################### .. GENERATED FROM PYTHON SOURCE LINES 324-369 .. code-block:: default def plot_trellis_with_segments(trellis, segments, transcript): # To plot trellis with path, we take advantage of 'nan' value trellis_with_path = trellis.clone() for i, seg in enumerate(segments): if seg.label != "|": trellis_with_path[seg.start + 1 : seg.end + 1, i + 1] = float("nan") fig, [ax1, ax2] = plt.subplots(2, 1, figsize=(16, 9.5)) ax1.set_title("Path, label and probability for each label") ax1.imshow(trellis_with_path.T, origin="lower") ax1.set_xticks([]) for i, seg in enumerate(segments): if seg.label != "|": ax1.annotate(seg.label, (seg.start + 0.7, i + 0.3), weight="bold") ax1.annotate(f"{seg.score:.2f}", (seg.start - 0.3, i + 4.3)) ax2.set_title("Label probability with and without repetation") xs, hs, ws = [], [], [] for seg in segments: if seg.label != "|": xs.append((seg.end + seg.start) / 2 + 0.4) hs.append(seg.score) ws.append(seg.end - seg.start) ax2.annotate(seg.label, (seg.start + 0.8, -0.07), weight="bold") ax2.bar(xs, hs, width=ws, color="gray", alpha=0.5, edgecolor="black") xs, hs = [], [] for p in path: label = transcript[p.token_index] if label != "|": xs.append(p.time_index + 1) hs.append(p.score) ax2.bar(xs, hs, width=0.5, alpha=0.5) ax2.axhline(0, color="black") ax2.set_xlim(ax1.get_xlim()) ax2.set_ylim(-0.1, 1.1) plot_trellis_with_segments(trellis, segments, transcript) plt.tight_layout() plt.show() .. image-sg:: /tutorials/images/sphx_glr_forced_alignment_tutorial_004.png :alt: Path, label and probability for each label, Label probability with and without repetation :srcset: /tutorials/images/sphx_glr_forced_alignment_tutorial_004.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 370-377 Looks good. Now let’s merge the words. The Wav2Vec2 model uses ``'|'`` as the word boundary, so we merge the segments before each occurance of ``'|'``. Then, finally, we segment the original audio into segmented audio and listen to them to see if the segmentation is correct. .. GENERATED FROM PYTHON SOURCE LINES 377-401 .. code-block:: default # Merge words def merge_words(segments, separator="|"): words = [] i1, i2 = 0, 0 while i1 < len(segments): if i2 >= len(segments) or segments[i2].label == separator: if i1 != i2: segs = segments[i1:i2] word = "".join([seg.label for seg in segs]) score = sum(seg.score * seg.length for seg in segs) / sum(seg.length for seg in segs) words.append(Segment(word, segments[i1].start, segments[i2 - 1].end, score)) i1 = i2 + 1 i2 = i1 else: i2 += 1 return words word_segments = merge_words(segments) for word in word_segments: print(word) .. rst-class:: sphx-glr-script-out .. code-block:: none I (0.78): [ 30, 34) HAD (0.84): [ 36, 43) THAT (0.52): [ 44, 52) CURIOSITY (0.89): [ 55, 91) BESIDE (0.94): [ 94, 115) ME (0.67): [ 117, 123) AT (0.66): [ 125, 128) THIS (0.70): [ 130, 138) MOMENT (0.97): [ 142, 154) .. GENERATED FROM PYTHON SOURCE LINES 402-404 Visualization ############################################################################### .. GENERATED FROM PYTHON SOURCE LINES 404-453 .. code-block:: default def plot_alignments(trellis, segments, word_segments, waveform): trellis_with_path = trellis.clone() for i, seg in enumerate(segments): if seg.label != "|": trellis_with_path[seg.start + 1 : seg.end + 1, i + 1] = float("nan") fig, [ax1, ax2] = plt.subplots(2, 1, figsize=(16, 9.5)) ax1.imshow(trellis_with_path[1:, 1:].T, origin="lower") ax1.set_xticks([]) ax1.set_yticks([]) for word in word_segments: ax1.axvline(word.start - 0.5) ax1.axvline(word.end - 0.5) for i, seg in enumerate(segments): if seg.label != "|": ax1.annotate(seg.label, (seg.start, i + 0.3)) ax1.annotate(f"{seg.score:.2f}", (seg.start, i + 4), fontsize=8) # The original waveform ratio = waveform.size(0) / (trellis.size(0) - 1) ax2.plot(waveform) for word in word_segments: x0 = ratio * word.start x1 = ratio * word.end ax2.axvspan(x0, x1, alpha=0.1, color="red") ax2.annotate(f"{word.score:.2f}", (x0, 0.8)) for seg in segments: if seg.label != "|": ax2.annotate(seg.label, (seg.start * ratio, 0.9)) xticks = ax2.get_xticks() plt.xticks(xticks, xticks / bundle.sample_rate) ax2.set_xlabel("time [second]") ax2.set_yticks([]) ax2.set_ylim(-1.0, 1.0) ax2.set_xlim(0, waveform.size(-1)) plot_alignments( trellis, segments, word_segments, waveform[0], ) plt.show() .. image-sg:: /tutorials/images/sphx_glr_forced_alignment_tutorial_005.png :alt: forced alignment tutorial :srcset: /tutorials/images/sphx_glr_forced_alignment_tutorial_005.png :class: sphx-glr-single-img .. GENERATED FROM PYTHON SOURCE LINES 455-469 .. code-block:: default # A trick to embed the resulting audio to the generated file. # `IPython.display.Audio` has to be the last call in a cell, # and there should be only one call par cell. def display_segment(i): ratio = waveform.size(1) / (trellis.size(0) - 1) word = word_segments[i] x0 = int(ratio * word.start) x1 = int(ratio * word.end) print(f"{word.label} ({word.score:.2f}): {x0 / bundle.sample_rate:.3f} - {x1 / bundle.sample_rate:.3f} sec") segment = waveform[:, x0:x1] return IPython.display.Audio(segment.numpy(), rate=bundle.sample_rate) .. GENERATED FROM PYTHON SOURCE LINES 471-477 .. code-block:: default # Generate the audio for each segment print(transcript) IPython.display.Audio(SPEECH_FILE) .. rst-class:: sphx-glr-script-out .. code-block:: none I|HAD|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 479-482 .. code-block:: default display_segment(0) .. rst-class:: sphx-glr-script-out .. code-block:: none I (0.78): 0.604 - 0.684 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 484-487 .. code-block:: default display_segment(1) .. rst-class:: sphx-glr-script-out .. code-block:: none HAD (0.84): 0.724 - 0.865 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 489-492 .. code-block:: default display_segment(2) .. rst-class:: sphx-glr-script-out .. code-block:: none THAT (0.52): 0.885 - 1.046 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 494-497 .. code-block:: default display_segment(3) .. rst-class:: sphx-glr-script-out .. code-block:: none CURIOSITY (0.89): 1.107 - 1.831 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 499-502 .. code-block:: default display_segment(4) .. rst-class:: sphx-glr-script-out .. code-block:: none BESIDE (0.94): 1.891 - 2.314 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 504-507 .. code-block:: default display_segment(5) .. rst-class:: sphx-glr-script-out .. code-block:: none ME (0.67): 2.354 - 2.474 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 509-512 .. code-block:: default display_segment(6) .. rst-class:: sphx-glr-script-out .. code-block:: none AT (0.66): 2.515 - 2.575 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 514-517 .. code-block:: default display_segment(7) .. rst-class:: sphx-glr-script-out .. code-block:: none THIS (0.70): 2.615 - 2.776 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 519-522 .. code-block:: default display_segment(8) .. rst-class:: sphx-glr-script-out .. code-block:: none MOMENT (0.97): 2.857 - 3.098 sec .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 523-529 Conclusion ---------- In this tutorial, we looked how to use torchaudio’s Wav2Vec2 model to perform CTC segmentation for forced alignment. .. rst-class:: sphx-glr-timing **Total running time of the script:** ( 0 minutes 2.589 seconds) .. _sphx_glr_download_tutorials_forced_alignment_tutorial.py: .. only:: html .. container:: sphx-glr-footer sphx-glr-footer-example .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: forced_alignment_tutorial.py ` .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: forced_alignment_tutorial.ipynb ` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_