Fine-Tuning VCTK Model destroys quality

Hey all,

I am attempting to continue fine-tuning an existing model. This is a first attempt into training voice models. My hope was to take a working model, fine-tune it and see the ongoing, improving results after continued updated epochs.

I understand this may be a premature evaluation but from a working Tacotron2 model to just after the first epoch it sounds unrecognizable as human speech. I expected a very fine, gradual improvement baselining somewhere near where the original model ended up. Instead, it’s quite a huge drop to the point you can’t understand it. So I am trying to understand where I misconfigured / setup my training. Below are my configs and starting points:

Model : Multi-Speaker-Tacotron2 DDC using source code based on commit 6cc464e. (I tried the advertised 2136433 commit but there were issues getting it to the model to align to the code so I reviewed the GIT log and used a downstream commit where fixes appeared to be merged. This seemed like a safe commit to attempt and it did not provide any model-based warnings or errors when run.

I pulled the model and configurations from the associated Colab Notebook.

The configuration I ran with is:

"github_branch":"* origin2_dev",
    "model": "Tacotron2",          // one of the model in models/  
    "run_name": "vctk-r=2-ddc",
    "run_description": "tacotron2 on vctl r=2 with ddc only without guided attention", 
    "mixed_precision": false,

        "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
        "win_length": 1024,      // stft window length in ms.
        "hop_length": 256,       // stft window hop-lengh in ms.
        "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
        "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.

        // Audio processing parameters
        "sample_rate": 48000,   // DATASET-RELATED: wav sample-rate. If different than the original data, it is resampled.
        "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
        "ref_level_db": 0,     // reference level db, theoretically 20db is the sound of air.

        // Silence trimming
        "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (false), TWEB (false), Nancy (true)
        "trim_db": 25,          // threshold for timming silence. Set this according to your dataset.

	// Griffin-Lim
	"power": 1.5,           // value to sharpen wav signals after GL algorithm.
	"griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.

        // MelSpectrogram parameters
        "num_mels": 80,         // size of the mel spec frame.
        "mel_fmin": 50.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
        "mel_fmax": 7600.0,     // maximum freq level for mel-spec. Tune for dataset!!
        "spec_gain": 1.0,         // scaler value appplied after log transform of spectrogram.

        // Normalization parameters
        "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
        "min_level_db": -100,   // lower bound for normalization
        "symmetric_norm": true, // move normalization to range [-1, 1]
        "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
        "clip_norm": true,      // clip normalized values into the range.
        //"stats_path": "/workspace/scale_stats.npy"    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by ''. If it is defined, mean-std based notmalization is used and other normalization params are ignored
        "stats_path": null    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by ''. If it is defined, mean-std based notmalization is used and other normalization params are ignored

    // if custom character set is not defined,
    // default set in is used
        "pad": "_",
        "eos": "&",
        "bos": "*",
        //"characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
        //"characters": "ABCDEFGHIJKLMçãàáâêéíóôõúûabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
        "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZÇÃÀÁÂÊÉÍÓÔÕÚÛabcdefghijklmnopqrstuvwxyzçãàáâêéíóôõúû!(),-.:;? ",
        "punctuations":"!'(),-.:;? ",
        "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ'̃' "
        "backend": "nccl",
        "url": "tcp:\/\/localhost:54322"

    "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.

    "batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
    "r": 2,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.  
    "gradual_training": [[0, 7, 32], [1, 5, 32], [100000, 3, 16], [250000, 2, 16], [500000, 1, 16]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.

    "loss_masking": false,       // enable / disable loss masking against the sequence padding.
    "decoder_loss_alpha": 0.5,  // decoder loss weight. If > 0, it is enabled
    "postnet_loss_alpha": 0.25, // postnet loss weight. If > 0, it is enabled
    "ga_alpha": 0.0,           // weight for guided attention loss. If > 0, guided attention is enabled.
    "decoder_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
    "postnet_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
    "decoder_ssim_alpha": 0.5,     // differential spectral loss weight. If > 0, it is enabled
    "postnet_ssim_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled

    "run_eval": true,
    "test_delay_epochs": 1,  //Until attention is aligned, testing only wastes computation time.
    "test_sentences_file": null,//"../../../datasets/BRSpeech-3-Speakers-Paper/BRSpeech-3-Speakers-Paper/TTS-Portuguese_Corpus/test_setences.txt",  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.

    "noam_schedule": false,        // use noam warmup and lr schedule.
    "grad_clip": 1.0,                // upper limit for gradients for clipping.
    "epochs": 1000,                // total number of epochs to train.
    "lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
    "wd": 0.000001,         // Weight decay weight.
    "warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
    "seq_len_norm": true,	   // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
    "memory_size": -1,              // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame. 
    "prenet_type": "original",     // "original" or "bn".
    "prenet_dropout": true,        // enable/disable dropout at prenet. 

    "attention_type": "original",  // 'original' or 'graves'
    "attention_heads": 4,          // number of attention heads (only for 'graves')
    "attention_norm": "softmax",   // softmax or sigmoid. Suggested to use softmax for Tacotron2 and sigmoid for Tacotron.
    "windowing": false,            // Enables attention windowing. Used only in eval mode.
    "use_forward_attn": false,      // if it uses forward attention. In general, it aligns faster.
    "forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
    "transition_agent": false,     // enable/disable transition agent of forward attention.
    "location_attn": true,        // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
    "bidirectional_decoder": false,  // use Use it, if attention does not work well with your dataset.
    "double_decoder_consistency": true,  // use DDC explained here
    "ddc_r": 4,                           // reduction rate for coarse decoder.
    "apex_amp_level": null, // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.

    // STOPNET
    "stopnet": true,               // Train stopnet predicting the end of synthesis. 
    "separate_stopnet": true,     // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.

    "print_step": 25,       // Number of steps to log traning on console.
    "tb_plot_step": 100,    // Number of steps to plot TB training figures.
    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
    "save_step": 250,      // Number of training steps expected to save traninpg stats and checkpoints.
    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
    "tb_model_param_stats": true,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging. 
    "text_cleaner": "english_cleaners",
    "enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
    "num_loader_workers": 8,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
    "num_val_loader_workers": 8,    // number of evaluation data loader processes.
    "batch_group_size": 0,  //Number of batches to shuffle after bucketing.
    "min_seq_len": 2,       // DATASET-RELATED: minimum text length to use in training
    "max_seq_len": 153,     // DATASET-RELATED: maximum text length

    // PATHS
    // "output_path": "/data5/rw/pit/keep/",      // DATASET-RELATED: output path for all training outputs.
    "output_path": "/data/output/LJSpeech/",
    "phoneme_cache_path": "/workspace/phonemes/",  // phoneme computation is slow, therefore, it caches results in the given folder.
    "use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
    "phoneme_language": "en-us",     // depending on your target language, pick one from

    // "speaker_encoder_config_path":"/root/speaker_encoder/LibriTTS-common-voice-voxceleb_angle_proto/config.json", // config.json for the speaker encoder
    // "speaker_encoder_checkpoint_path": "/root/speaker_encoder/LibriTTS-common-voice-voxceleb_angle_proto/320k.pth.tar", // Speaker Encoder Checkpoint full path
    "use_speaker_embedding": true,     // use speaker embedding to enable multi-speaker learning.
    "use_external_speaker_embedding_file": true, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: /1806.04558
    "external_speaker_embedding_file": "/workspace/train/speakers.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: /1806.04558
    "use_gst": false,       // TACOTRON ONLY: use global style tokens

    "gst":	{			                // gst parameter if gst is enabled
        "gst_style_input": null,        // Condition the style input either on a 
                                        // -> wave file [path to wave] or 
                                        // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}  
                                        // with the dictionary being len(dict) <= len(gst_style_tokens).
        "disable_gst": false, // if true its force GST predict zero matrix.
        "gst_use_speaker_embedding": true, // if true pass speaker embedding in attention input GST.
        "gst_embedding_dim": 128,       
        "gst_num_heads": 4,
        "gst_style_tokens": 5
        // DATASETS
        "datasets":   // List of datasets. They all merged and they get different s$
                "name": "vctk",
                "path": "/data/VCTK/VCTK-Corpus",
                // para o teste no vctk eu escolhi 1 locutor de cada ACCEPTS, 
                "meta_file_train": ["p225", "p234", "p238", "p245", "p248", "p261", "p294", "p302", "p326", "p335", "p347"], // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
                "meta_file_val": null


As stated above, this configuration runs as expected without warnings or errors. But when I review the test audio via the Tensorboard, it’s indistinguishable from straight noise. (Attached) (528.5 KB) (rename to test_audio.wav without unzipping)

Further, the supporting images seem to indicate a “start over” scenario.

Finally, like I said, I didn’t see any warnings or errors on the application startup. I include the first portion of the application log here:

python /workspace/TTS/TTS/bin/ --script /workspace/TTS/TTS/bin/ --config_path=/workspace/train/config.json --continue_path=/workspace/output/ --restore_path=/workspace/train/tts_model.pth.tar'
2021-02-18 16:14:28.576959: I tensorflow/stream_executor/platform/default/] Successfully opened dynamic library
2021-02-18 16:14:28.592577: I tensorflow/stream_executor/platform/default/] Successfully opened dynamic library
 > Using CUDA:  True
 > Number of GPUs:  2
 > Training continues for /workspace/output/
 > Setting up Audio Processor...
 | > sample_rate:48000
 | > num_mels:80
 | > min_level_db:-100
 | > frame_shift_ms:None
 | > frame_length_ms:None
 | > ref_level_db:0
 | > fft_size:1024
 | > power:1.5
 | > preemphasis:0.0
 | > griffin_lim_iters:60
 | > signal_norm:True
 | > symmetric_norm:True
 | > mel_fmin:50.0
 | > mel_fmax:7600.0
 | > spec_gain:1.0
 | > stft_pad_mode:reflect
 | > max_norm:4.0
 | > clip_norm:True
 | > do_trim_silence:True
 | > trim_db:25
 | > do_sound_norm:False
 | > stats_path:None
 | > hop_length:256
 | > win_length:1024
 | > Found 39490 files in /data/VCTK/VCTK-Corpus
Training with 96 speakers: VCTK_p226, VCTK_p227, VCTK_p228, VCTK_p229, VCTK_p230, VCTK_p231, VCTK_p232, VCTK_p233, VCTK_p236, VCTK_p237, VCTK_p239, VCTK_p240, VCTK_p241, VCTK_p243, VCTK_p244, VCTK_p246, VCTK_p247, VCTK_p249, VCTK_p250, VCTK_p251, VCTK_p252, VCTK_p253, VCTK_p254, VCTK_p255, VCTK_p256, VCTK_p257, VCTK_p258, VCTK_p259, VCTK_p260, VCTK_p262, VCTK_p263, VCTK_p264, VCTK_p265, VCTK_p266, VCTK_p267, VCTK_p268, VCTK_p269, VCTK_p270, VCTK_p271, VCTK_p272, VCTK_p273, VCTK_p274, VCTK_p275, VCTK_p276, VCTK_p277, VCTK_p278, VCTK_p279, VCTK_p280, VCTK_p281, VCTK_p282, VCTK_p283, VCTK_p284, VCTK_p285, VCTK_p286, VCTK_p287, VCTK_p288, VCTK_p292, VCTK_p293, VCTK_p295, VCTK_p297, VCTK_p298, VCTK_p299, VCTK_p300, VCTK_p301, VCTK_p303, VCTK_p304, VCTK_p305, VCTK_p307, VCTK_p308, VCTK_p310, VCTK_p311, VCTK_p312, VCTK_p313, VCTK_p314, VCTK_p316, VCTK_p317, VCTK_p318, VCTK_p323, VCTK_p329, VCTK_p330, VCTK_p333, VCTK_p334, VCTK_p336, VCTK_p339, VCTK_p340, VCTK_p341, VCTK_p343, VCTK_p345, VCTK_p351, VCTK_p360, VCTK_p361, VCTK_p362, VCTK_p363, VCTK_p364, VCTK_p374, VCTK_p376
 > Using model: Tacotron2
 > Model restored from step 220000

 > Model has 51314996 parameters

 > EPOCH: 0/1000

 > Number of output frames: 2

 > DataLoader initialization
 | > Use phonemes: True
   | > phoneme language: en-us
 | > Number of instances : 39096
 | > Max length sequence: 181
 | > Min length sequence: 9
 | > Avg length sequence: 39.60463986085533
 | > Num. instances discarded by max-min (max=153, min=2) seq limits: 184
 | > Batch group size: 0.

 > TRAINING (2021-02-18 16:16:17) 

   --> STEP: 24/1216 -- GLOBAL_STEP: 220025
     | > decoder_loss: 4.02636  (4.76018)
     | > postnet_loss: 3.88876  (4.63819)
     | > stopnet_loss: 0.36340  (0.33130)
     | > decoder_coarse_loss: 5.08758  (5.75460)
     | > decoder_ddc_loss: 0.01690  (0.01789)
     | > decoder_diff_spec_loss: 0.05701  (0.08949)
     | > postnet_diff_spec_loss: 0.04774  (0.07405)
     | > decoder_ssim_loss: 0.49228  (0.51219)
     | > postnet_ssim_loss: 0.49113  (0.51146)
     | > loss: 5.76943  (6.75253)
     | > align_error: 0.62557  (0.60317)
     | > avg_spec_length: 281.0
     | > avg_text_length: 15.2
     | > step_time: 3.8240
     | > loader_time: 0.01
     | > current_lr: 0.0001

Any help or insight into what I’m misunderstanding would be greatly appreciated. Thanks much for the library and the hard work!

Your spectrograms look weird, there were others with such problem, try setting “spec_gain” to 20.0, it’s the default when not using a stats path.

Also did I understand it right that you only trained for one epoch? It will require more epochs, even when finetuning to get decent results.

I think you don’t have the scale stats (audio.stats_path)computed and used for your run. Or as @sanjaesc suggested use spec_gain 20 but then you need to train your own vocoder.

So I have scale_stats set to null due to the inline comments stating not to use it with multi-speaker data sets. I’ll get that back in place and if audio still seems off, I’ll go with @sanjaesc’s suggestion.

Thank you!

Thank you! I was wondering if that was the case or not… After letting it run all night it still sounds horrible but that’s secondary to the scale_stats nullification… Fixing that now and will let her run the weekend.

Thank you!

Follow-up status and question:

After using the compute_statistics script against my dataset, pointing the config file at the resultant scale_stats.npy and letting that run for the last 4 days, the results are much better. Voices and utterances are generally clear but there are 2 issues that I would like to understand how to address:

  1. During inference, upwards of 95% of the attempts are met with the Decoder stopped with 'max_decoder_steps error. I have read up on this error and understand that simply put, something in the target text is either too complex or basically confusing the application. Game over. Understanding that, but also understanding that phrases such as, “This is a test” are being met with that error, I look to the model. How can we train it to be more robust? Or, am I misunderstanding something here? I am using the same config.json that we used for training to run the inference. Is there a better inference configuration that we should be considering? (Same config as above minus we are pointing to a valid scale_stats.npy now).

  2. The training itself seems to have stagnated. The voice quality of tests today closely resembles those of the tests that I had Sunday morning. Minimal progress, if any. Noting we are training this against Librit tts-clean-100. How can I help “nudge” the training back into gear? (Audio still is tinny / robotty, has periods of garbage noise, etc.)

Thank you … please direct to existing doc’s if there are some that I have not been mentioned. (Sorry, lots to read as I get up to speed.)