app.py

app.py

@@ -0,0 +1,79 @@
+import gradio as gr
+import librosa
+import numpy as np
+import torch
+
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+
+
+processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+model = SpeechT5ForTextToSpeech.from_pretrained("tejas1206/speecht5_tts_technical_en")
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+
+
+speaker_embeddings = {
+    "BDL": "speaker/cmu_us_bdl_arctic-wav-arctic_a0009.npy",
+    "CLB": "speaker/cmu_us_clb_arctic-wav-arctic_a0144.npy",
+    "KSP": "speaker/cmu_us_ksp_arctic-wav-arctic_b0087.npy",
+    "RMS": "speaker/cmu_us_rms_arctic-wav-arctic_b0353.npy",
+    "SLT": "speaker/cmu_us_slt_arctic-wav-arctic_a0508.npy",
+}
+
+
+def predict(text, speaker):
+    if len(text.strip()) == 0:
+        return (16000, np.zeros(0).astype(np.int16))
+
+    inputs = processor(text=text, return_tensors="pt")
+
+    # limit input length
+    input_ids = inputs["input_ids"]
+    input_ids = input_ids[..., :model.config.max_text_positions]
+
+    if speaker == "Surprise Me!":
+        # load one of the provided speaker embeddings at random
+        idx = np.random.randint(len(speaker_embeddings))
+        key = list(speaker_embeddings.keys())[idx]
+        speaker_embedding = np.load(speaker_embeddings[key])
+
+        # randomly shuffle the elements
+        np.random.shuffle(speaker_embedding)
+
+        # randomly flip half the values
+        x = (np.random.rand(512) >= 0.5) * 1.0
+        x[x == 0] = -1.0
+        speaker_embedding *= x
+
+        #speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
+    else:
+        speaker_embedding = np.load(speaker_embeddings[speaker[:3]])
+
+    speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
+
+    speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
+
+    speech = (speech.numpy() * 32767).astype(np.int16)
+    return (16000, speech)
+
+
+title = "Text-to-Speech App using SpeechT5"
+
+gr.Interface(
+    fn=predict,
+    inputs=[
+        gr.Text(label="Input Text"),
+        gr.Radio(label="Speaker", choices=[
+            "BDL (male)",
+            "CLB (female)",
+            "KSP (male)",
+            "RMS (male)",
+            "SLT (female)",
+            "Surprise Me!"
+        ],
+        value="BDL (male)"),
+    ],
+    outputs=[
+        gr.Audio(label="Generated Speech", type="numpy"),
+    ],
+    title=title,
+).launch()

requirements.txt

@@ -0,0 +1,8 @@
+gradio==5.1.0
+torch==2.4.0
+git+https://github.com/huggingface/transformers.git
+soundfile==0.12.1
+sentencepiece==0.2.0
+samplerate
+librosa
+resampy

speaker/cmu_us_awb_arctic-wav-arctic_a0002.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5db7a684ab490f21cec1628e00d461a184e369fe4eafb1ee441a796faf4ab6ae
+size 2176

speaker/cmu_us_bdl_arctic-wav-arctic_a0009.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:215326eae3a428af8934c385fbe043b36c72849ca17d1d013adeb189e6bd6962
+size 2176

speaker/cmu_us_clb_arctic-wav-arctic_a0144.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf67b36c47edfb1851466a1dff081b436bc6809b5ebc12811d9df0c0d0f28d0e
+size 2176

speaker/cmu_us_ksp_arctic-wav-arctic_b0087.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6c5c2a38c2e400179019c560a74c4322f4ee13beda22ee601807545edee283e
+size 2176

speaker/cmu_us_rms_arctic-wav-arctic_b0353.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a49dac3e9c3a71a4dbca4c364233c7915ae6e0cb71b2ceaed97296231b95cb50
+size 2176

speaker/cmu_us_slt_arctic-wav-arctic_a0508.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f71ffadda3f3a4de079740a0b34963824dc644d9d5442283bd0a2b0d4f44ff0b
+size 2176