Garbled(�) non-ASCII characters in tokens[*].text in JSON output when --dtw used.

[MUCZ]

Description:

When using the main cli program to transcribe audio containing non-ASCII characters and outputting detailed JSON --output-json-full with features DTW enabled, the transcription.tokens[*].text field sometimes contains garbled characters (represented as � ). However, the main transcription.text field appears correct.

For example:

"text": "えー、こんばんは。岸高野、岸大輔です。本日は心を入れ替えまして、TBSラジオのスタジオから収録しております。どういうことなんだよ、それ。心を入れ替えてすることじゃねえだろ。本来そっちなんだよ。",
			"tokens": [
				{
					"text": "[_BEG_]",
					"timestamps": {
						"from": "00:00:00,000",
						"to": "00:00:00,000"
					},
					"offsets": {
						"from": 0,
						"to": 0
					},
					"id": 50365,
					"p": 0.813801,
					"t_dtw": -1
				},
				{
					"text": "えー",
					"timestamps": {
						"from": "00:00:00,000",
						"to": "00:00:00,470"
					},
					"offsets": {
						"from": 0,
						"to": 470
					},
					"id": 36697,
					"p": 0.0519727,
					"t_dtw": 158
				},
				{
					"text": "、",
					"timestamps": {
						"from": "00:00:00,470",
						"to": "00:00:00,700"
					},
					"offsets": {
						"from": 470,
						"to": 700
					},
					"id": 1231,
					"p": 0.824364,
					"t_dtw": 196
				},
				{
					"text": "こん",
					"timestamps": {
						"from": "00:00:00,930",
						"to": "00:00:01,170"
					},
					"offsets": {
						"from": 930,
						"to": 1170
					},
					"id": 31172,
					"p": 0.990851,
					"t_dtw": 210
				},
				{
					"text": "ば",
					"timestamps": {
						"from": "00:00:01,400",
						"to": "00:00:01,400"
					},
					"offsets": {
						"from": 1400,
						"to": 1400
					},
					"id": 13349,
					"p": 0.998904,
					"t_dtw": 220
				},
				{
					"text": "ん",
					"timestamps": {
						"from": "00:00:01,420",
						"to": "00:00:01,630"
					},
					"offsets": {
						"from": 1420,
						"to": 1630
					},
					"id": 3225,
					"p": 0.99433,
					"t_dtw": 228
				},
				{
					"text": "は",
					"timestamps": {
						"from": "00:00:01,630",
						"to": "00:00:01,860"
					},
					"offsets": {
						"from": 1630,
						"to": 1860
					},
					"id": 3065,
					"p": 0.993292,
					"t_dtw": 236
				},
				{
					"text": "。",
					"timestamps": {
						"from": "00:00:01,890",
						"to": "00:00:02,090"
					},
					"offsets": {
						"from": 1890,
						"to": 2090
					},
					"id": 1543,
					"p": 0.638299,
					"t_dtw": 280
				},
				{
					"text": "�", # E5 B2
					"timestamps": {
						"from": "00:00:02,090",
						"to": "00:00:02,240"
					},
					"offsets": {
						"from": 2090,
						"to": 2240
					},
					"id": 23182,
					"p": 0.940747,
					"t_dtw": 318
				},
				{
					"text": "�", # B8
					"timestamps": {
						"from": "00:00:02,240",
						"to": "00:00:02,310"
					},
					"offsets": {
						"from": 2240,
						"to": 2310
					},
					"id": 116,
					"p": 0.999584,
					"t_dtw": 326
				},
				{
					"text": "高",
					"timestamps": {
						"from": "00:00:02,310",
						"to": "00:00:02,350"
					},
					"offsets": {
						"from": 2310,
						"to": 2350
					},
					"id": 12979,
					"p": 0.375837,
					"t_dtw": 342
				},
				{
					"text": "野",
					"timestamps": {
						"from": "00:00:02,540",
						"to": "00:00:02,770"
					},
					"offsets": {
						"from": 2540,
						"to": 2770
					},
					"id": 37178,
					"p": 0.900068,
					"t_dtw": 358
				},
....

Noted that the single character 岸 is shown correctly in text field but splitted into 2 tokens, the utf-8 encoding for 岸 is E5 B2 B8, corresponding to text field of the 2 tokens: E5 B2 and B8.

imo the root cause for this is that the tokenizer sometimes splits single non-ASCII characters into multiple tokens.

I found a workaround for decoding this using python, simply by concatenating consecutive tokens that is garbled. I'm not sure if this will/should be solved from the whisper.cpp side.

def merge_token_by_utf8(inputs: list[Token]) -> list[Token]:
    out = []
    i = 0
    while i < len(inputs):
        if '\\' in inputs[i].text:
            to_merge = [inputs[i]]
            j = 0
            merged = None
            while j < len(inputs):
                j+=1
                if '\\' in inputs[i+j].text:
                    to_merge.append(inputs[i+j])
                    try:
                        merged = TokenList.merge_tokens(to_merge)
                        matches = list(re.finditer(r'(\\x[0-9a-fA-F]{2})', merged.text))
                        prefix,suffix='',''
                        if len(matches)>0:
                            if matches[0].start() == 0:
                                split_point = matches[-1].end() 
                                merged.text = merged.text[:split_point]
                                suffix = merged.text[split_point:]
                            else:
                                split_point = matches[0].start()
                                prefix =  merged.text[:split_point]
                                merged.text = merged.text[split_point:]
                        merged.text = merged.text.encode('latin1').decode('unicode_escape').encode('latin1').decode('utf-8')
                        merged.text = prefix + merged.text + suffix
                        break
                    except Exception as e:
                        continue
                else:
                    raise Exception(f"failed to merge {inputs[i+j].text}")
            print(f'merged: [{merged.text}]')
            out.append(merged)
            i+=j
        else:
            out.append(inputs[i])
        i+=1
    return out 
            
def read() -> List[Token]:
    raw = open(json_file_path, errors='backslashreplace').read().replace('\\','\\\\')
    raw = json.loads(raw)
    return merge_jap_token_by_utf8(raw)