Upload IndicF5 MNN models and sample audio

.gitattributes

@@ -33,3 +33,17 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+indicf5_dit_fp16.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_dit_fp16_int8wq.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_dit_fp32.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_dit_int8wq.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_fp16.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_fp16_int8wq.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_fp32.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_fp32_dynamic.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_int8wq.mnn filter=lfs diff=lfs merge=lfs -text
+indicf5_vocos_int8wq_dynamic.mnn filter=lfs diff=lfs merge=lfs -text
+samples/ta_medium_mnn_vocos_fp32.wav filter=lfs diff=lfs merge=lfs -text
+samples/ta_medium_mnn_vocos_int8wq.wav filter=lfs diff=lfs merge=lfs -text
+samples/ta_short_mnn_vocos_fp32.wav filter=lfs diff=lfs merge=lfs -text
+samples/ta_short_mnn_vocos_int8wq.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,52 @@
+---
+license: mit
+library_name: mnn
+tags:
+  - text-to-speech
+  - indic
+  - mnn
+  - vocos
+  - f5-tts
+  - indian-languages
+  - quantization
+pipeline_tag: text-to-speech
+base_model: ai4bharat/IndicF5
+---
+
+# IndicF5 MNN
+
+MNN exports of [ai4bharat/IndicF5](https://huggingface.co/ai4bharat/IndicF5) (DiT + Vocos), including fp16 and int8 weight-quantized variants.
+
+## Files
+
+### DiT
+
+| File | Approx. size |
+|------|----------------|
+| `indicf5_dit_fp32.mnn` | ~1.3 GB |
+| `indicf5_dit_fp16.mnn` | ~646 MB |
+| `indicf5_dit_int8wq.mnn` | ~347 MB |
+| `indicf5_dit_fp16_int8wq.mnn` | ~646 MB |
+
+### Vocos
+
+| File | Notes |
+|------|--------|
+| `indicf5_vocos_fp32.mnn` | static |
+| `indicf5_vocos_fp32_dynamic.mnn` | variable mel frames (recommended for audio) |
+| `indicf5_vocos_fp16.mnn` | |
+| `indicf5_vocos_int8wq.mnn` | static |
+| `indicf5_vocos_int8wq_dynamic.mnn` | variable mel frames |
+| `indicf5_vocos_fp16_int8wq.mnn` | |
+
+Sample WAVs (fp32 / int8wq Vocos decode) are under `samples/`. Bench numbers: `quant_benchmark.json`.
+
+## Notes
+
+- Prefer `*_dynamic.mnn` Vocos models for full-length audio (call `resizeTensor` with a **tuple** shape, then `resizeSession`).
+- Path used for samples: PyTorch mel → MNN Vocos.
+- Still not realtime on Apple M1 CPU for full DiT ODE + Vocos.
+
+## Attribution
+
+Derived from [ai4bharat/IndicF5](https://huggingface.co/ai4bharat/IndicF5). Follow the base model license/terms for redistribution and use.

indicf5_dit_fp16.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2ec9100d3e0182b90bb42c2b6bda0b7bf3d9a716133a219aa8f57b2bd24f70
+size 676662616

indicf5_dit_fp16_int8wq.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2ec9100d3e0182b90bb42c2b6bda0b7bf3d9a716133a219aa8f57b2bd24f70
+size 676662616

indicf5_dit_fp32.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94ea8f0607222a4d0f5b5a13139e651c68652e2e8627b20050b9f850f7b8a248
+size 1350402900

indicf5_dit_int8wq.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d19cdb4ef19b6209b0945c900f285031c2c36882256ec10b4afeb0f45f460dd
+size 346515768

indicf5_vocos_fp16.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:770529207742282a441b9d7520c0807ad4584073f1d2c3c70d0cf8384df308b6
+size 29342284

indicf5_vocos_fp16_int8wq.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:770529207742282a441b9d7520c0807ad4584073f1d2c3c70d0cf8384df308b6
+size 29342284

indicf5_vocos_fp32.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a0abedc44b86643d2b292baaab5604811bb766f8865837a00e7934dced19fa3
+size 58506108

indicf5_vocos_fp32_dynamic.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:574450cfc77f08ca76c08f876b34a5b241ddfa8c4d2f8f7d93b53d53fcfdbf61
+size 62580312

indicf5_vocos_int8wq.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cc5ed902a84404402256a1be8e58439f2986462b1f1e2c3575e552e839c81f9
+size 15053088

indicf5_vocos_int8wq_dynamic.mnn

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f944f50d6a9af95220ae3ff4fce8594ab9374f2005aa5157bed11eb69d01729
+size 19127292

quant_benchmark.json

@@ -0,0 +1,231 @@
+{
+  "conversions": {
+    "mnn_vocos_fp16": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_vocos_fp16.mnn",
+    "mnn_vocos_fp16_mb": 29.34,
+    "mnn_vocos_int8wq": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_vocos_int8wq.mnn",
+    "mnn_vocos_int8wq_mb": 15.05,
+    "mnn_vocos_fp16_int8wq": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_vocos_fp16_int8wq.mnn",
+    "mnn_vocos_fp16_int8wq_mb": 29.34,
+    "mnn_dit_fp16": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_dit_fp16.mnn",
+    "mnn_dit_fp16_mb": 676.66,
+    "mnn_dit_int8wq": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_dit_int8wq.mnn",
+    "mnn_dit_int8wq_mb": 346.52,
+    "mnn_dit_fp16_int8wq": "/Users/abu/qefro/indicf5-benchmark/mnn/indicf5_dit_fp16_int8wq.mnn",
+    "mnn_dit_fp16_int8wq_mb": 676.66,
+    "litert_vocos_fp16": "/Users/abu/qefro/indicf5-benchmark/litert/vocos/indicf5_vocos_fp16.tflite",
+    "litert_vocos_fp16_mb": 29.29,
+    "litert_vocos_drq": "/Users/abu/qefro/indicf5-benchmark/litert/vocos_quant/indicf5_vocos_dynamic_range_quant.tflite",
+    "litert_vocos_drq_mb": 57.23,
+    "litert_dit_drq": null
+  },
+  "benchmarks": [
+    {
+      "name": "mnn_vocos_fp32",
+      "ok": true,
+      "mean_s": 0.01844218359997285,
+      "std_s": 0.00017029317445372552,
+      "min_s": 0.0182243340000241,
+      "max_s": 0.01870145899988529,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 58.51,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_dit_fp32",
+      "ok": true,
+      "mean_s": 0.3114288501999454,
+      "std_s": 0.03828318255141011,
+      "min_s": 0.24100124999995387,
+      "max_s": 0.34994941700006166,
+      "out_shape": [
+        1,
+        128,
+        100
+      ],
+      "size_mb": 1350.4,
+      "format": "mnn"
+    },
+    {
+      "name": "litert_vocos_fp32",
+      "ok": true,
+      "mean_s": 0.023510274599948387,
+      "std_s": 0.00024219493729862353,
+      "min_s": 0.023173624999799358,
+      "max_s": 0.02386070799980189,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 58.52,
+      "format": "litert"
+    },
+    {
+      "name": "mnn_vocos_fp16",
+      "ok": true,
+      "mean_s": 0.01884079139995265,
+      "std_s": 0.0003826578183155768,
+      "min_s": 0.018479916999922352,
+      "max_s": 0.01943533300004674,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 29.34,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_vocos_int8wq",
+      "ok": true,
+      "mean_s": 0.018562100000008287,
+      "std_s": 0.0001635790577244952,
+      "min_s": 0.01843074999987948,
+      "max_s": 0.01886475000014798,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 15.05,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_vocos_fp16_int8wq",
+      "ok": true,
+      "mean_s": 0.021306441400111,
+      "std_s": 0.0028035859396329985,
+      "min_s": 0.01785941600019214,
+      "max_s": 0.024554208000154176,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 29.34,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_dit_fp16",
+      "ok": true,
+      "mean_s": 0.48576437519996035,
+      "std_s": 0.1993106179657564,
+      "min_s": 0.26494529199999306,
+      "max_s": 0.7715467499999704,
+      "out_shape": [
+        1,
+        128,
+        100
+      ],
+      "size_mb": 676.66,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_dit_int8wq",
+      "ok": true,
+      "mean_s": 0.2773376831999485,
+      "std_s": 0.03149785049287247,
+      "min_s": 0.23587449999990895,
+      "max_s": 0.3237904999998591,
+      "out_shape": [
+        1,
+        128,
+        100
+      ],
+      "size_mb": 346.52,
+      "format": "mnn"
+    },
+    {
+      "name": "mnn_dit_fp16_int8wq",
+      "ok": true,
+      "mean_s": 0.2746680413999911,
+      "std_s": 0.03585345627868485,
+      "min_s": 0.22788883299995177,
+      "max_s": 0.33330849999993006,
+      "out_shape": [
+        1,
+        128,
+        100
+      ],
+      "size_mb": 676.66,
+      "format": "mnn"
+    },
+    {
+      "name": "litert_vocos_fp16",
+      "ok": false,
+      "error": "tflite/kernels/conv.cc:360 input_type == kTfLiteFloat32 || input_type == kTfLiteUInt8 || input_type == kTfLiteInt8 || input_type == kTfLiteInt16 was not true.Node number 3 (CONV_2D) failed to prepare.",
+      "size_mb": 29.29
+    },
+    {
+      "name": "litert_vocos_drq",
+      "ok": true,
+      "mean_s": 0.02443846679998387,
+      "std_s": 0.00026067359768839086,
+      "min_s": 0.024193834000016068,
+      "max_s": 0.024878499999886117,
+      "out_shape": [
+        1,
+        1,
+        32512
+      ],
+      "size_mb": 57.23,
+      "format": "litert"
+    },
+    {
+      "name": "litert_dit_drq",
+      "ok": false,
+      "error": "missing"
+    }
+  ],
+  "quantization": [
+    {
+      "model": "mnn_vocos_fp16",
+      "ok": true,
+      "size_mb": 29.34
+    },
+    {
+      "model": "mnn_vocos_int8wq",
+      "ok": true,
+      "size_mb": 15.05
+    },
+    {
+      "model": "mnn_vocos_fp16_int8wq",
+      "ok": true,
+      "size_mb": 29.34
+    },
+    {
+      "model": "mnn_dit_fp16",
+      "ok": true,
+      "size_mb": 676.66
+    },
+    {
+      "model": "mnn_dit_int8wq",
+      "ok": true,
+      "size_mb": 346.52
+    },
+    {
+      "model": "mnn_dit_fp16_int8wq",
+      "ok": true,
+      "size_mb": 676.66
+    },
+    {
+      "model": "litert_vocos_fp16",
+      "ok": true,
+      "size_mb": 29.29
+    },
+    {
+      "model": "litert_vocos_drq",
+      "ok": true,
+      "size_mb": 57.23
+    },
+    {
+      "model": "litert_dit_drq",
+      "ok": false,
+      "error": "Skipped: DiT LiteRT float is not runnable (ONNX_CONV); DRQ reconvert would be ~1.3GB+ and same blocker."
+    }
+  ]
+}

samples/ta_medium_mnn_vocos_fp32.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3825a0ac3403de806610f2baeedc6e2611a676a9bf8202bb1840820ae31b1d5
+size 374828

samples/ta_medium_mnn_vocos_int8wq.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ac896dac715b49f04f8d0bc6733dc2db9c3176e3839ed7a970fadb4b072dd45
+size 374828

samples/ta_short_mnn_vocos_fp32.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6d3022ad723c825464beb8e72a9962bf3a5b85b52469c93fd5520b5d45a924b
+size 157740

samples/ta_short_mnn_vocos_int8wq.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4649003425f2b363dc72629d5348074f9eee47fbcfd799922b8f5288eddb41e4
+size 157740