XTTS: add inference_stream_text (slightly friendlier for text-streaming)

TTS/tts/models/xtts.py

@@ -209,6 +209,8 @@ def __init__(self, config: Coqpit):
         self.decoder_checkpoint = self.args.decoder_checkpoint  # TODO: check if this is even needed
         self.models_dir = config.model_dir
         self.gpt_batch_size = self.args.gpt_batch_size
+        self._stream_text_holder = []
+        self._stream_generator = None
 
         self.tokenizer = VoiceBpeTokenizer()
         self.gpt = None
@@ -632,64 +634,140 @@ def inference_stream(
         length_scale = 1.0 / max(speed, 0.05)
         gpt_cond_latent = gpt_cond_latent.to(self.device)
         speaker_embedding = speaker_embedding.to(self.device)
-        if enable_text_splitting:
-            text = split_sentence(text, language, self.tokenizer.char_limits[language])
-        else:
-            text = [text]
+        text_streaming = (text is None)
 
-        for sent in text:
-            sent = sent.strip().lower()
-            text_tokens = torch.IntTensor(self.tokenizer.encode(sent, lang=language)).unsqueeze(0).to(self.device)
+        while True:
+            if text_streaming:
+                yield None
+                if len(self._stream_text_holder) == 0:
+                    return
+                text, enable_text_splitting = self._stream_text_holder
 
-            assert (
-                text_tokens.shape[-1] < self.args.gpt_max_text_tokens
-            ), " ❗ XTTS can only generate text with a maximum of 400 tokens."
+            if enable_text_splitting:
+                text = split_sentence(text, language, self.tokenizer.char_limits[language])
+            else:
+                text = [text]
 
-            fake_inputs = self.gpt.compute_embeddings(
-                gpt_cond_latent.to(self.device),
-                text_tokens,
-            )
-            gpt_generator = self.gpt.get_generator(
-                fake_inputs=fake_inputs,
-                top_k=top_k,
-                top_p=top_p,
-                temperature=temperature,
-                do_sample=do_sample,
-                num_beams=1,
-                num_return_sequences=1,
-                length_penalty=float(length_penalty),
-                repetition_penalty=float(repetition_penalty),
-                output_attentions=False,
-                output_hidden_states=True,
-                **hf_generate_kwargs,
-            )
+            for sent in text:
+                sent = sent.strip().lower()
+                text_tokens = torch.IntTensor(self.tokenizer.encode(sent, lang=language)).unsqueeze(0).to(self.device)
+
+                assert (
+                    text_tokens.shape[-1] < self.args.gpt_max_text_tokens
+                ), " ❗ XTTS can only generate text with a maximum of 400 tokens."
+
+                fake_inputs = self.gpt.compute_embeddings(
+                    gpt_cond_latent.to(self.device),
+                    text_tokens,
+                )
+                gpt_generator = self.gpt.get_generator(
+                    fake_inputs=fake_inputs,
+                    top_k=top_k,
+                    top_p=top_p,
+                    temperature=temperature,
+                    do_sample=do_sample,
+                    num_beams=1,
+                    num_return_sequences=1,
+                    length_penalty=float(length_penalty),
+                    repetition_penalty=float(repetition_penalty),
+                    output_attentions=False,
+                    output_hidden_states=True,
+                    **hf_generate_kwargs,
+                )
+
+                last_tokens = []
+                all_latents = []
+                wav_gen_prev = None
+                wav_overlap = None
+                is_end = False
+
+                while not is_end:
+                    try:
+                        x, latent = next(gpt_generator)
+                        last_tokens += [x]
+                        all_latents += [latent]
+                    except StopIteration:
+                        is_end = True
+
+                    if is_end or (stream_chunk_size > 0 and len(last_tokens) >= stream_chunk_size):
+                        gpt_latents = torch.cat(all_latents, dim=0)[None, :]
+                        if length_scale != 1.0:
+                            gpt_latents = F.interpolate(
+                                gpt_latents.transpose(1, 2), scale_factor=length_scale, mode="linear"
+                            ).transpose(1, 2)
+                        wav_gen = self.hifigan_decoder(gpt_latents, g=speaker_embedding.to(self.device))
+                        wav_chunk, wav_gen_prev, wav_overlap = self.handle_chunks(
+                            wav_gen.squeeze(), wav_gen_prev, wav_overlap, overlap_wav_len
+                        )
+                        last_tokens = []
+                        yield wav_chunk
+
+            if not text_streaming:
+                return
+
+    def inference_stream_text(
+        self,
+        language,
+        gpt_cond_latent,
+        speaker_embedding,
+        # Streaming
+        stream_chunk_size=20,
+        overlap_wav_len=1024,
+        # GPT inference
+        temperature=0.75,
+        length_penalty=1.0,
+        repetition_penalty=10.0,
+        top_k=50,
+        top_p=0.85,
+        do_sample=True,
+        speed=1.0,
+        **hf_generate_kwargs,
+    ):
+        if self._stream_generator is not None:
+            raise Exception('Inference text-streaming already in progress. '
+                            'Did you forget to call inference_finalize_text?')
+
+        # Arguments `text` and `enable_text_splitting` given through holder
+        self._stream_text_holder = [None, None]
+        self._stream_generator = self.inference_stream(
+            None,
+            language,
+            gpt_cond_latent,
+            speaker_embedding,
+            stream_chunk_size=stream_chunk_size,
+            overlap_wav_len=overlap_wav_len,
+            temperature=temperature,
+            length_penalty=length_penalty,
+            repetition_penalty=repetition_penalty,
+            top_k=top_k,
+            top_p=top_p,
+            do_sample=do_sample,
+            speed=speed,
+            **hf_generate_kwargs,
+        )
 
-            last_tokens = []
-            all_latents = []
-            wav_gen_prev = None
-            wav_overlap = None
-            is_end = False
-
-            while not is_end:
-                try:
-                    x, latent = next(gpt_generator)
-                    last_tokens += [x]
-                    all_latents += [latent]
-                except StopIteration:
-                    is_end = True
-
-                if is_end or (stream_chunk_size > 0 and len(last_tokens) >= stream_chunk_size):
-                    gpt_latents = torch.cat(all_latents, dim=0)[None, :]
-                    if length_scale != 1.0:
-                        gpt_latents = F.interpolate(
-                            gpt_latents.transpose(1, 2), scale_factor=length_scale, mode="linear"
-                        ).transpose(1, 2)
-                    wav_gen = self.hifigan_decoder(gpt_latents, g=speaker_embedding.to(self.device))
-                    wav_chunk, wav_gen_prev, wav_overlap = self.handle_chunks(
-                        wav_gen.squeeze(), wav_gen_prev, wav_overlap, overlap_wav_len
-                    )
-                    last_tokens = []
-                    yield wav_chunk
+        # Start the generator and return it
+        _ = next(self._stream_generator)
+        return self._stream_generator
+
+    def inference_add_text(self, text: str, enable_text_splitting=False):
+        if self._stream_generator is None:
+            raise Exception('Inference text-streaming not started. '
+                            'Please call inference_stream_text first')
+        self._stream_text_holder[0] = text
+        self._stream_text_holder[1] = enable_text_splitting
+
+    def inference_finalize_text(self):
+        if self._stream_generator is None:
+            raise Exception('Inference text-streaming was not started '
+                            '(start with inference_stream_text)')
+        # Finalize and reset the generator
+        self._stream_text_holder.clear()
+        try:
+            _ = next(self._stream_generator)
+        except StopIteration:
+            pass
+        self._stream_generator = None
 
     def forward(self):
         raise NotImplementedError(

docs/source/models/xtts.md

@@ -220,7 +220,7 @@ torchaudio.save("xtts.wav", torch.tensor(out["wav"]).unsqueeze(0), 24000)
 ```
 
 
-##### Streaming manually
+##### Streaming inference
 
 Here the goal is to stream the audio as it is being generated. This is useful for real-time applications.
 Streaming inference is typically slower than regular inference, but it allows to get a first chunk of audio faster.
@@ -253,16 +253,50 @@ chunks = model.inference_stream(
     speaker_embedding
 )
 
-wav_chuncks = []
+wav_chunks = []
 for i, chunk in enumerate(chunks):
     if i == 0:
         print(f"Time to first chunck: {time.time() - t0}")
     print(f"Received chunk {i} of audio length {chunk.shape[-1]}")
-    wav_chuncks.append(chunk)
-wav = torch.cat(wav_chuncks, dim=0)
+    wav_chunks.append(chunk)
+wav = torch.cat(wav_chunks, dim=0)
 torchaudio.save("xtts_streaming.wav", wav.squeeze().unsqueeze(0).cpu(), 24000)
 ```
 
+If you also need to do text-streaming you can use `inference_stream_text`, like so:
+
+```python
+# ...same setup as before
+
+def text_streaming_generator():
+    yield "It took me quite a long time to develop a voice and now that I have it I am not going to be silent."
+    yield "Having discovered not just one, but many voices, I will champion each."
+
+print("Inference with text streaming...")
+
+text_gen = text_streaming_generator()
+inf_gen = model.inference_stream_text(
+    "en",
+    gpt_cond_latent,
+    speaker_embedding
+)
+
+wav_chunks = []
+for text in text_gen:
+    # Add text progressively
+    model.inference_add_text(text, enable_text_splitting=True)
+    for chunk in enumerate(inf_gen):
+        if chunk is None:
+            break # all chunks generated for the current text
+        print(f"Received chunk {len(wav_chunks)} of audio length {chunk.shape[-1]}")
+        wav_chunks.append(chunk)
+
+# Call finalize to discard the inference generator
+model.inference_finalize_text()
+
+wav = torch.cat(wav_chunks, dim=0)
+torchaudio.save("xtts_streaming_text.wav", wav.squeeze().unsqueeze(0).cpu(), 24000)
+```
 
 ### Training