[codex] Add CLI outputs, language detection, punctuation removal, and streaming demos

README.md

@@ -42,14 +42,14 @@ pip install -U dataoceanai-dolphin
 
 Alternatively, it can also be installed from the source:
 ```shell
-pip install git+https://github.com/SpeechOceanTech/Dolphin.git 
+pip install git+https://github.com/DataoceanAI/Dolphin.git
 ```
 
 ## Available Models and Languages
 
 ### Models
 
-There are 8 models in Dolphin, and 6 of them are available now. See details in [Dolphin](https://arxiv.org/abs/2503.20212) and [Dolphin-CN-Dialect](https://arxiv.org/abs/2605.08961).
+Available Dolphin models are listed below. See details in [Dolphin](https://arxiv.org/abs/2503.20212) and [Dolphin-CN-Dialect](https://arxiv.org/abs/2605.08961).
 
 |  Model  | Parameters |Publicly Available |
 |:------:|:----------:|:------------------:|
@@ -58,8 +58,8 @@ There are 8 models in Dolphin, and 6 of them are available now. See details in [
 | medium |   0.9 B    |            |
 | large  |   1.7B   |             |
 | [base.cn](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-base) | 0.1 B | ✅ |
-| [base.cn.streaming](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-small-prompt) | 0.1 B | ✅ |
-| [small.cn](https://modelscope.cn/models/DataoceanAI/dolphi-cn-dialect-small) | 0.4 B | ✅ |
+| [base.cn.streaming](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-base-streaming) | 0.1 B | ✅ |
+| [small.cn](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-small) | 0.4 B | ✅ |
 | [small.cn.streaming](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-small-streaming) | 0.4 B | ✅ |
 | [small.cn.prompt](https://modelscope.cn/models/DataoceanAI/dolphin-cn-dialect-small-prompt) | 0.4 B | ✅ |
 
@@ -75,12 +75,27 @@ Dolphin supports 40 Eastern languages and 22 Chinese dialects. For a complete li
 # default model:small
 dolphin audio.wav
 
+# Write plain text output to a file
+dolphin audio.wav --output result.txt
+
+# Write structured output with metadata
+dolphin audio.wav --output result.json --output_format json
+
+# Write subtitle output
+dolphin audio.wav --output result.srt --output_format srt
+
 # Download model and specify the model path
 dolphin audio.wav --model small.cn
 
 # Specify language and region
 dolphin audio.wav --model small.cn --lang_sym "zh" --region_sym "CN"
 
+# Detect language and region only. This uses the Dolphin ASR model's built-in
+# language identification head; this package does not provide a separate
+# lightweight LID-only model.
+dolphin audio.wav --model small.cn --task detect_language
+dolphin long_audio.wav --model small.cn --task detect_language --lid_duration 30
+
 # Specify the hotwords file with Encoder-biased method
 dolphin audio.wav --model small.cn --hotword_list_path hotwords.txt --use_deep_biasing true
 
@@ -90,8 +105,43 @@ dolphin audio.wav --model small.cn.prompt --hotword_list_path hotwords.txt --use
 # predict word timestamp
 dolphin audio.wav --model small.cn.prompt --word_timestamp true
 
+# Remove punctuation from transcription text
+dolphin audio.wav --model small.cn --remove_punctuation true
+
 ```
 
+### Experimental streaming demo
+
+For Chinese dialect streaming models, the repository provides an experimental
+cache-level streaming demo. It drives `forward_encoder_chunk` with encoder
+caches and prints CTC partial results as each chunk is decoded. `--chunk_size`
+controls the encoder streaming chunk size. CTC endpointing is enabled by
+default, so a long silence or long utterance automatically finalizes the
+current segment:
+
+```shell
+python examples/streaming_demo.py audio.wav --model small.cn.streaming --device cuda --chunk_size 16 --final_rescore attention
+```
+
+For timestamped partial lines or CPU smoke tests:
+
+```shell
+python examples/streaming_demo.py audio.wav --model base.cn.streaming --device cpu --chunk_size 16 --emit line --max_chunks 2
+```
+
+To stream from your microphone, install the optional recorder dependency and
+run:
+
+```shell
+python -m pip install sounddevice
+python examples/microphone_streaming_demo.py --model small.cn.streaming --device cuda --chunk_size 16 --final_rescore attention
+```
+
+Endpoint defaults follow common CTC streaming behavior: 5s silence before any
+decoded text, 1s silence after decoded text, or 20s maximum utterance length.
+Use `--disable_endpoint` to turn this off, or tune
+`--endpoint_rule2_min_trailing_silence_ms` for faster/slower segment finals.
+
 ### Python usage
 
 ```python
@@ -104,10 +154,18 @@ model = dolphin.load_model(model_name, device="cuda")
 result = transcribe(model, 'audio.wav')
 print(result.text)
 
+# Detect language and region only
+language, region = dolphin.detect_language(model, 'audio.wav')
+print(language, region)
+
 # Specify language
 result = transcribe(model, 'audio.wav', lang_sym="zh")
 print(result.text)
 
+# Remove punctuation from transcription text
+result = transcribe(model, 'audio.wav', remove_punctuation=True)
+print(result.text_nospecial)
+
 # Specify language and region and encoder-biased hotwords
 result = transcribe(model, 'audio.wav', lang_sym="zh", region_sym="CN", hotwords=['诺香丹青牌科研胶囊'], use_deep_biasing=True, use_two_stage_filter=True)
 print(result.text)
@@ -129,7 +187,6 @@ Thanks to the following excellent open-source works:
 - [Espnet](https://github.com/espnet/espnet)
 - [Wenet](https://github.com/wenet-e2e/wenet)
 - [FunASR](https://github.com/modelscope/FunASR)
-- [FireRedASR2S](https://github.com/FireRedTeam/FireRedASR2S)
 
 ## License
 

dolphin/__init__.py

@@ -1,6 +1,6 @@
 # encoding: utf8
 
 from .audio import load_audio
-from .transcribe import load_model, transcribe
+from .transcribe import detect_language, load_model, transcribe
 from .hotword import HotwordEncoder, apply_deep_biasing, two_stage_filtering
 from .version import __version__

dolphin/model.py

@@ -960,11 +960,16 @@ def forward(
 
             return self.forward_attention(v, scores, mask), new_cache
         else:
-            # NOTE(Mddct): we need mask bias, not boolean mask
-            assert mask.dtype != torch.bool
-            mask = mask.unsqueeze(1)
-            # matrix_bd as a mask bias
-            mask = (matrix_bd + mask) / math.sqrt(self.d_k)
+            if mask.size(-1) > 0:
+                # NOTE(Mddct): SDPA needs an attention bias here so the
+                # relative position logits can share the same attn_mask.
+                if mask.dtype == torch.bool:
+                    mask = mask_to_bias(mask, query.dtype)
+                mask = mask.unsqueeze(1)
+                mask = matrix_bd + mask
+            else:
+                mask = matrix_bd
+            mask = mask / math.sqrt(self.d_k)
             output = torch.nn.functional.scaled_dot_product_attention(
                 q_with_bias_u,
                 k,

dolphin/model_registry.py

@@ -17,7 +17,7 @@
         "sha256": "62e4c11fe1e0e42bd34e444172c5a05e792c4b5a03750f794fa3206fc0649cd7"
     },
     "small.cn": {
-        "model_id": "DataoceanAI/dolphi-cn-dialect-small",
+        "model_id": "DataoceanAI/dolphin-cn-dialect-small",
         "sha256": "1cee2b8d2133cabb36567625a832d4033569e27eaf5f98df9be1139ec6068bbb",
     },
     "small.cn.streaming": {